Imaging Pearls ❯ December 2025
|
-- OR -- |
|
3D and Workflow
- Introduction: Early detection of cancer improves survival following diagnosis. However, routine screening is limited to a few cancer types. Multicancer early detection (MCED) tests could revolutionize cancer screening by simultaneously detecting multiple cancer types. This study evaluates the potential impact of an MCED test on stage shift in the US general population.
Conclusion: MCED testing has the potential to substantially reduce late stage cancer diagnosis, improve outcomes across multiple cancer types and address critical gap in screening.
The impact of multicancer early detection tests on cancer stage shift: A 10-year microsimulation model.
Chhatwal J, Xiao J, ElHabr AK, Tyson C, Cao X, Raoof S, Fendrick AM, Ozbay AB, Limburg P, Beer TM, Briggs A, Deshmukh AA.
Cancer. 2025 Nov 15;131(22):e70075. - Over the 10-year horizon, supplemental testing with an annual MCED test resulted in a 10% increase (3364 vs 3068 cases per 100,000) in Stage I diagnoses, 20% increase (2491 vs 2079) in Stage II diagnoses, and 34% increase (1896 vs 1414) in Stage III diagnoses, relative to the SoC alone; in contrast, Stage IV diagnoses decreased by 45% (1159 vs 2108) The cumulative number of diagnoses was 8669 under the SoC, and 8910 when supplemented by MCED testing, equating to a modest increase of 2.8% (241 per 100,000). Of these 241 additional diagnoses, 82 were made in individuals who died from non–cancer-related causes under the SoC after their counterfactual time of MCED diagnosis, and 159 were in individuals who were eventually diagnosed under the SoC after the first 10 years. Figure 1B depicts the flow of individuals from their stage at diagnosis under the SoC to their stage at diagnosis when SoC is supplemented with MCED testing.
The impact of multicancer early detection tests on cancer stage shift: A 10-year microsimulation model.
Chhatwal J, Xiao J, ElHabr AK, Tyson C, Cao X, Raoof S, Fendrick AM, Ozbay AB, Limburg P, Beer TM, Briggs A, Deshmukh AA.
Cancer. 2025 Nov 15;131(22):e70075. - Our study shows that MCED testing has the potential to substantially reduce Stage IV cancer incidence, particularly for cancer types that lack routine screening programs. Although further research is needed to validate these findings in real-world settings, our results suggest that MCED testing could transform cancer diagnosis and improve patient outcomes across a broad range of cancer types.
The impact of multicancer early detection tests on cancer stage shift: A 10-year microsimulation model.
Chhatwal J, Xiao J, ElHabr AK, Tyson C, Cao X, Raoof S, Fendrick AM, Ozbay AB, Limburg P, Beer TM, Briggs A, Deshmukh AA.
Cancer. 2025 Nov 15;131(22):e70075.
- Has radiology already irreversibly abdicated its position as a field that can lead transformative patient care academic endeavors? There are numerous examples of the academic pursuits of radiology being subverted to benefit other specialties that may be viewed as being more aligned with leadership’s view of how the academic mission should be deployed [5]. Perhaps we are at a crossroads where we are required us to think about how we speak to the leadership of the health systems to reflect the tripartite mission. Perhaps it is our actual mission to be clear about what we should be doing. In these challenging times, remaining silent may not be an option. We need to be clear what is important and what is not.
The Academic Mission Starts, or Ends, at the Top.
Fishman EK, Lee DJ, Chu LC, Rowe SP.
J Am Coll Radiol. 2025 Oct 2:S1546-1440(25)00574-5. - At the heart of every academic medical department lies the incontrovertible paradox of the need to generate the revenue to maintain the department and the inevitable overlying bureaucracy, while at the same time pursuing the teaching, research, innovation, dissemination, and other elements that are crucial to the tripartite mission. Such considerations are perhaps particularly felt in a field such as radiology, in which an individual radiologist can produce a tremendous amount of revenue for a department and an institution, such that any time spent on the academic mission may become viewed as a loss of potential revenue.
The Academic Mission Starts, or Ends, at the Top.
Fishman EK, Lee DJ, Chu LC, Rowe SP.
J Am Coll Radiol. 2025 Oct 2:S1546-1440(25)00574-5. - It is imperative that radiology leaders be selected for their commitment to the academic mission, as opposed to how well they generate RVUs to be funneled to the rest of the health care system. In choosing leadership, whether at a Fortune 500 company or in academic medicine, it is important to look at a candidate’s beliefs, plans, and past performance, with a lens on their prior accomplishments. In the not too distant past, one might have asked the question whether that person was a “triple threat,” having excelled at- research, teaching, and clinical work. However, in the current era, an academic radiology chair needs to be a “quadruple threat,” not only subsuming all aspects of the traditional tripartite mission but also being financially savvy to ensure that funds flow both to the school of medicine and to mission-aligned projects within the department.
The Academic Mission Starts, or Ends, at the Top.
Fishman EK, Lee DJ, Chu LC, Rowe SP.
J Am Coll Radiol. 2025 Oct 2:S1546-1440(25)00574-5. - The recent paper by Subha Ghosh, MD, MBA, titled “The Silent Retreat: Quiet Quitting in the Academic Radiology Workforce” eloquently laid out a series of concerns that have become increasingly relevant within the radiology community in recent years. In an era when scientific discovery is commonplace— whether it be CT, MRI, molecular imaging and targeted radionuclide therapy, or generative artificial intelligence—it is difficult to believe that we have reached a point of academic stagnation. If anything, discoveries are continuing at a prodigious rate. From ostensible meritocracy to clinically skewed incentives, Dr Ghosh outlines many relevant factors that create an environment ripe for “quiet quitting.” These comments are sad but true, and they pose grave risk to the spirit of inquiry that defines academic radiology as we know it. Although Dr Ghosh accurately suggests that “our institutions must evolve,” we ask, who will lead this effort?
The Role of Leadership in Quiet Quitting.
Fishman EK, Rowe SP, Smith CW.
J Am Coll Radiol. 2025 Oct 2:S1546-1440(25)00572-1. - Drastic change is indeed necessary, but as Albert Einstein suggested, our problems cannot be solved with the same thinking that created them. Medicine is, in great part, now run by businesspeople who view physicians much as we view equipment or a- bottle of iodinated contrast. Each can be reduced to numbers, and physicians often become cogs in the overall machinery of a hospital. Radiologists are told to keep reading and generate relative value units (RVUs), while administrators reach their quotas and clinical output is maximized. A study by Eschelman et al found a stark inverse relationship between RVU output and articles, presentations, and abstracts produced. -
The Role of Leadership in Quiet Quitting.
Fishman EK, Rowe SP, Smith CW.
J Am Coll Radiol. 2025 Oct 2:S1546-1440(25)00572-1. - In the context of the current institutional model that incentivizes maximal clinical productivity and minimal academic output, if we want radiologists—and physicians at large—to not only provide care but advance it, we need to find forward-thinking leadership that will rekindle the fizzling embers of academic radiology, encouraging radiologists to grow and succeed beyond the reading room. Such leaders may be hard to find, as they must balance the traditional tripartite mission with the need to still generate the RVUs that help run the health care enterprise.
The Role of Leadership in Quiet Quitting.
Fishman EK, Rowe SP, Smith CW.
J Am Coll Radiol. 2025 Oct 2:S1546-1440(25)00572-1.
- Introduction: Three-dimensional (3D) reconstruction transforms cross-sectional medical images into interactive anatomical models, interpretable on an LCD screen, in augmented reality or via 3D printing. Although certain benefits have been established in liver surgery, its use in pancreatic surgery remains limited. This update outlines the applications of 3D visualization in pancreatic surgery, ranging from surgical planning to teaching.
Results: The analysis of these studies suggests that 3D reconstruction, in comparison to cross-sectional imaging, could improve preoperative evaluation, by facilitating the detection of anatomical variations, the assessment of resection margins, and the prediction of morbidity and mortality according to tumor volume and residual pancreatic parenchyma. 3D imaging could also improve intraoperative safety, with some series reporting a 50% reduction of blood loss and a 25% reduction in operating time. 3D reconstruction is also a promising tool for teaching surgical anatomy, particularly through 3D printing.
Conclusion: 3D reconstruction could improve outcomes of pancreatic surgery but requires robust comparative studies before becoming a standard evidence-based practice.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - Conversely, the use of 3D reconstructions in pancreatic surgery is still limited, and the scientific literature on the subject remains scarce. However, pancreatic surgery is also notorious for its complexity: resectability is determined by anatomic vascular relationships, the risk of bleeding is high,postoperative complications are frequent, and the prognosis of pancreatic disease, particularly malignancies, remains generally bleak with a low 5-year survival rate. These characteristics make 3D reconstructions a particularly relevant field, both to improve surgical planning and anticipate technical difficulties as well as to refine resectability criteria.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - 3D reconstruction could offer a substantial advantage over 2D imaging in the preoperative evaluation of resection margins, especially that of arterial resection, which isa key element in the surgical strategy of pancreatic tumor surgery. In a retrospective study of 105 patients, Griser et al. showed that the diagnostic performance (area under the curve of the receiver-operator characteristic curve) forthe assessment of arterial invasion was statistically significantly better than that of 2D. These results were confirmed by Fang et al., who, thanks to an improvement in sensitivity, specificity, and positive and negative predictive values provided by 3D reconstruction, proposed a new classification of tumor resectability based on these three-dimensional models. One of the hypotheses put forward toexplain this superiority is that 2D imaging tends to under-estimate the true tumor volume and that there is acorrelation between tumor volume, TNM stage and the risk of recurrence.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - 3D reconstruction of the pancreas is currently based on 2D images (CT or MRI), the quality of which depends on the thickness of the slices and the stability of the patient. Man-ual or semi-automated reconstructions are time-consuming and require expertise. Outsourcing processing by specialized private companies or the development of automated tools,particularly through deep learning, could improve the accuracy and reduce processing time. Medico-economic studies comparing the different solutions should be conducted to move from a frenzy for innovation to evidence-based practices.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - To date, the traditional open laparotomy approach remains preferable in pancreatic surgery in many settings, particularly in the case of a locally advanced tumor or a complex surgical procedure. Regarding PD, which isthe most common intervention, randomized trials comparing laparoscopic versus the open route have not formally demonstrated a clear benefit of the minimally invasive approach in terms of morbidity and mortality or oncologi-cal outcomes . Similarly, the robotic approach, although innovative, still needs to prove itself in high-level methodology studies before being generalized.In this context, AR could represent a real game-changer for minimally invasive approaches. Indeed, the possibility of projecting 3D reconstructions in real time directly on the screen during laparoscopic or robotic-assisted surgery could compensate for the absence of tactile and visual cues that are specific to open surgery. This technological integration, which is difficult to transpose to conventional surgery,would strengthen the safety and precision of gestures in minimally invasive procedures.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011
- Small pancreatic ductal adenocarcinomas (PDACs) generally have a more favorable prognosis; however, up to 40% may be missed on CT scan. Cinematic rendering (CR) is a novel three-dimensional post-processing technique that accentuates subtle textural differences and holds the potential to improve detection of small tumors. It utilizes a complex global lighting model generating photorealistic images with enhanced visualization of shadows and texture based on volumetric scalar fields (Hounsfield units). While CR has been studied for pre-surgical planning and vascular anatomy assessment, its role in detecting small PDACs remains underexplored. This pictorial essay highlights the potential of CR in detecting small PDACs and differentiating surface textures, while also addressing its limitations and future directions.
- Cinematic rendering (CR) is a newer 3D rendering technique, first described for PDAC assessment in 2018, that creates photorealistic images using a global lighting model. This approach improves anatomical visualization and depth perception. CR enhances the visibility of subtle textural differences, improving tumor conspicuity compared to conventional 2D reconstruction, 3D VR or MIP techniques. Unlike standard VR, which relies on a single light source, CR employs a more complex global lighting model incorporating multiple light sources. This approach enables more accurate representation of shadows, textures, and enhanced depth of field in the images.
- A central component of this approach is the transfer function, which maps scalar intensity values to material-specific properties such as color and opacity. These mappings allow different tissue types to be differentiated based on their density profiles and enable the selective enhancement or suppression of anatomical structures. The transfer function thereby implicitly defines "surfaces" within the volume as regions of high opacity gradients or significant radiometric contribution.
- Surface-like features emerge where the accumulated opacity reaches a perceptually significant threshold, often coinciding with sharp transitions in the scalar field. At these locations, a local gradient is used to approximate the surface normal, enabling shading via physically based lighting models. The decision to terminate, scatter, or continue a ray is governed by Monte Carlo sampling. Rays encountering highly opaque or emissive regions may be absorbed or contribute significantly to image synthesis through direct or indirect illumination. In contrast, rays traversing low-opacity regions are likely to proceed further into the volume, accumulating color and opacity until a termination criterion is met. Optimizing these factors may result in lesion detection which may be missed on routine CT post-processing.
- Pancreatic ductal adenocarcinomas (PDACs) are often detected at an unresectable stage, leading to high rates of morbidity and mortality.However, patients with small PDACs (≤2 cm) have demonstrated better overall prognosis. PDACs are primarily evaluated and staged using a three-phase intravenous (IV) contrast-enhanced pancreas protocol computed tomography (CT) scan, which is also the key modality for determining local resectability. The overall sensitivity of CT for PDAC detection is about 90 %, but it drops to 63–77 % for PDAC ≤2 cm in size. Studies suggest that up to 40 % of small PDACs may go undetected on CT scans, highlighting the considerable challenges associated with identifying these tumors using conventional imaging methods, and representing a significant missed opportunity for potential curative surgery.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - In the context of medical imaging, path tracing applied to volumetric data offers physically based rendering for the visualization of anatomical structures. Rather than relying on explicitly segmented surfaces, this technique reconstructs the appearance of tissue boundaries and internal features by interpreting volumetric scalar fields — e.g., Hounsfield units in CT scans — as continuous spatial distributions of optical properties. A central component of this approach is the transfer function, which maps scalar intensity values to material-specific properties such as color and opacity. These mappings allow different tissue types to be differentiated based on their density profiles and enable the selective enhancement or suppression of anatomical structures. The transfer function thereby implicitly defines "surfaces" within the volume as regions of high opacity gradients or significant radiometric contribution.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - Path tracing operates by casting rays into the volume and stochastically integrating the radiative transfer equation along each path. At discrete intervals along a ray’s trajectory, the local scalar field is sampled, and the corresponding optical properties are retrieved from the transfer function. Since medical volumes are stored on discrete grids, interpolation — typically trilinear — is employed to estimate scalar values between voxel centers. This interpolation smooths transitions but can also introduce intermediate values in regions with high gradients, such as tissue interfaces. As a result, the optical properties in these transition zones may be significantly altered, potentially softening the visual appearance of otherwise sharp boundaries or affecting the perceived thickness and translucency of structures.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - Cinematic rendering offers an opportunity to improve the detection of small PDACs through its enhanced, photorealistic visualization. This technique highlights subtle textural differences between pancreatic tumors and the surrounding normal parenchyma, making it easier to identify small lesions that might otherwise be missed. CR can enhance the understanding of the spatial relationship between a mass and surrounding structures, thereby increasing diagnostic confidence In our practice, we have developed a series of presets for different organs or clinical applications, including eight specific presets designed for the pancreas. However, these presets can be manipulated in real time to accentuate the optical surface properties for improved tumor detection for specific cases.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - As discussed above, the ability of CR to enhance surface differences and internal architecture offers the potential for improved diagnostic accuracy. Fig. 3 depicts a pancreatic head tumor that is visualized on both CT and cinematic rendering. The CR images accentuate the tumor’s surface texture, and contrast adjustments provide improved visualization of its internal architecture, as shown in images C and D of Fig. 3. Similarly, Fig. 4 demonstrates how CR enhances visualization of tumor density, texture, and vascular anatomy in an incidentally detected PDAC.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - As illustrated above, cinematic rendering has the capacity to enhance diagnostic accuracy in PDAC detection, particularly helpful for small lesions that may be iso-attenuating or too subtle to be identified by conventional imaging. Its interactive display settings can be tailored to the anatomy of interest, optimizing visualization of pathological processes. This can be utilized to differentiate PDAC from mimicking pancreatic lesions, including pancreatitis, mass forming pancreatitis, pancreatic neuroendocrine tumor (PNET), solid pseudopapillary neoplasm (SPN) or metastasis from a primary tumor. The capability to modify display settings can emphasize structures with higher Hounsfield units, such as PNETs. The improved depth perception enables sharper visualization of the internal architecture of cystic pancreatic neoplasms, including fine septations and mural nodularity. This improved detail may aid in distinguishing between several types of cystic neoplasms . CR has been shown valuable in visualizing the mixed cystic-solid components and vascular spatial relationships in cases of SPN, including those with rupture. Furthermore, it assists in differentiating vessel stretching caused by a large SPN, from true vessel involvement seen in PDAC, thereby enhancing diagnostic confidence.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - The most significant barrier is the complex algorithm utilized by CR that requires dedicated software along with additional computational power, substantial network bandwidth, and specialized training; all of which may not be routinely available in most clinical settings. The post-processing involved in interactive rendering to tailor display parameters for each clinical indication and pathology demands both high expertise and real-time calculations for each manipulation, further requiring longer processing times. Preset selection and parameter optimization are crucial for accurate representation of the anatomy and pathology of interest. This requires experience with the CR software to optimize parameters accurately for lesion identification. Incorrect adjustment of these settings can either hinder visualization of critical findings or lead to diagnostic errors by ‘creating’ new lesions that may not be truly present, particularly due to shadowing effects inherent to the lighting model. This may lead to overcall of clinically insignificant lesions leading to increased healthcare costs for the patients. Although CR significantly enhances photorealism, its clinical utility remains theoretical, and it is still unclear whether this improvement leads to greater diagnostic accuracy.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - It is anticipated that the parameter adjustments for a particular pathology can be standardized and potentially automated in the future.Currently, rendering each case with CR requires approximately 5–7 min when performed by an experienced radiologist, whereas less experienced radiologists may require additional time. However, the integration of artificial intelligence has the potential to streamline this process and significantly reduce rendering times.AI can potentially automate preset selection and optimization of parameters based on both patient related factors and the organ of interest to enhance lesion detection and avoid human biases. Furthermore, with the advancement of radiomics and machine learning, CR images could serve as a valuable resource for extracting prognostic information from high-order imaging features.Integration of CR technology with augmented reality devices like Microsoft’s HoloLens offers new opportunities for immersive surgical planning and better inter-provider communication, enabling real-time visualization and interaction with radiological images.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5.
Cardiac
- Purpose: To evaluate the impact of a fully automated, multitask deep learning (DL) algorithm on interreader agreement of coronary artery disease (CAD) detection and stenosis classification using coronary CT angiography (CCTA).
Materials and Methods: This retrospective study included CCTA examinations (n = 623 patients) performed for clinical indications on CT systems from multiple vendors between January 2010 and December 2019. An expert reader (reader 1) analyzed all CCTA scans manually and with artificial intelligence (AI)–assisted reading at the lesion, coronary segment, and patient levels using the CAD Reporting and Data System (CAD-RADS). The AI algorithm detected, quantified, and classified coronary lesions. Interreader agreement was evaluated using a second expert reader (reader 2), who analyzed a randomly selected subset of 274 patients. CAD-RADS scores from radiologist reports (reader 3) were available for 362 patients. In a subgroup of 30 patients with disagreements, R2 also interpreted the cases using AI assistance. Agreement between readings, with and without AI, was assessed using Spearman correlation,and logistic regression and mixed models evaluated the impact of AI-assisted reading on CAD-RADS classification.
Conclusion: AI-assisted reading using a DL algorithm significantly improved interreader agreement for CAD-RADS classification at CCTA.
Impact of Deep Learning–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 - ■ Artificial intelligence (AI)–assisted reading of coronary CT angiography improved interreader agreement between two expert readers for Coronary Artery Disease Reporting and Data System (CADRADS) classification, with Spearman correlation coefficient increasing from 0.899 to 0.949 (P < .001).
■ In cases in which reader 1 and the AI disagreed, interreader agreement between reader 1 and reader 2 was low with manual readings (ρ = 0.688) but significantly improved when both readers used AI-assisted reading (ρ = 0.975; P < .001).
■ AI-assisted reading of coronary CT angiography improved the agreement between an expert reader and the existing radiologist reports from multiple radiologists for CAD-RADS classification, with the Spearman correlation coefficient increasing from 0.889 to 0.938 (P < .001).
Impact of Deep Learning–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 
Impact of Deep Learning–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563- Finally, the role of AI-assisted reading in CAD-RADS reporting is not limited to standardization and variability reduction. In our study, AI helped reader 1 identify 11 patients with at least one plaque (CAD-RADS ≥ 0) that were labeled as CAD-RADS 0 during manual reading. The detection of small, noncalcified, easy-to-miss plaques is important because the presence of a plaque, even a small one, can prompt clinicians to start preventive therapy . Moreover, detecting early signs of CAD at CCTA is important, especially for prognostic reasons. In the PROMISE (Prospective Multicenter Imaging Study for Evaluation of Chest Pain) trial, a large prospective trial of CCTA in patients with stable chest pain and suspected CAD, CAD-RADS showed a greater prognostic value than all other traditional scores tested. Because CCTA is progressively assuming the role of a gatekeeper for unnecessary invasive coronary angiography, one of the main purposes of AI introduction to cardiovascular imaging is to reduce the number of false negatives.
ng–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 - In conclusion, the use of a DL algorithm for AI-assisted reading significantly improved interreader agreement in CADRADS classification of CCTA images. Future studies should test AI-assisted reading on larger, more robust datasets with consistent and overlapping expert annotations. They should also assess its clinical impact, real-time performance across diverse settings, and potential role in training less experienced readers. Correlation with invasive coronary angiography may further support its diagnostic value.
Impact of Deep Learning–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563
Chest
- Purpose: This study investigates the frequency, progression, and clinical implications of pancreatic findings on chest low-dose computed tomography (LDCT) scans performed for lung cancer screening.
Conclusion: Incidental pancreatic findings were uncommon (0.9%) and included calcifications, atrophy/fatty infiltration, cysts, ductal dilatation, and masses. These findings do not by themselves indicate pancreatic cancer but warrant documentation and, when suspicious, dedicated pancreatic imaging. Radiologist scrutiny could improve detection accuracy, indicating the potential of a LDCT lung cancer screening program for detecting and monitoring pancreatic lesions.
Pancreatic findings in participants in a program of low-dose computed tomography screening for lung cancer
Gros, Louisa,b; Yip, Rowenaa; Zhu, Yeqinga; Li, Pengfeia; Paksashvili, Natelaa; Sun, Qia; Yankelevitz, David F.a; Henschke, Claudia I.a
European Journal of Cancer Prevention ():10.1097/CEJ.0000000000000997, November 18, 2025 - Out of 9467 participants, 90 (0.9%) had pancreatic findings, mostly male (54.4%), median age 64.7, with smoking (92.2%), alcohol use (41.1%), and diabetes (22%). Of these, 60 (66.7%) were detected on baseline LDCT, primarily as calcifications (73.3%), atrophy/fatty infiltration (18.3%), and duct dilatation (5%). Of the 90 participants, 27 underwent only baseline LDCT. Among the remaining 63, 33 had pancreatic findings on baseline scans, 27 of whom (81.8%) showed consistent findings on follow-up, and 30 developed pancreatic findings during surveillance. Rereview of the baseline scans showed that 68 participants (75.6%) had findings, including eight missed earlier. More cases of atrophy/fatty infiltration and other findings were detected compared to the original report, with calcifications remaining predominant (50 participants). Similar patterns were observed during the rereview of the latest LDCT scans. Two participants with detected lesions underwent biopsy, diagnosing a serous cystadenoma and pancreatic adenocarcinoma. The latter succumbed to pancreatic cancer.
Pancreatic findings in participants in a program of low-dose computed tomography screening for lung cancer
Gros, Louisa,b; Yip, Rowenaa; Zhu, Yeqinga; Li, Pengfeia; Paksashvili, Natelaa; Sun, Qia; Yankelevitz, David F.a; Henschke, Claudia I.a
European Journal of Cancer Prevention ():10.1097/CEJ.0000000000000997, November 18, 2025
Colon
- It is important to note that most cases of enterocolitis are self-limiting and do not routinely require cross-sectional imaging. Indications for CT arise in patients with severe abdominal pain, peritoneal signs, systemic toxicity, or sepsis, as well as in those with underlying comorbidities or immunosuppression. Imaging is also justified in cases of diagnostic uncertainty, atypical presentations, failure to improve with conservative management, or when complications such as perforation, ischemia, or toxic megacolon are suspected.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - Inflammatory bowel disease (IBD) is a non-infectious cause of chronic relapsing–remitting gastrointestinal inflammation that occurs due to the combination of genetic, host, and environmental factors. IBD has a rising worldwide prevalence affecting 2.4–3.1 million people in the United States [6]. Crohn’s disease and ulcerative colitis constitute the two principal subtypes of inflammatory bowel disease (IBD), both commonly presenting with abdominal pain, diarrhea, and weight loss, with or without gastrointestinal bleeding in the form of melena or hematochezia. Acute gastroenteritis may serve as the initial clinical manifestation in patients with previously undiagnosed IBD, or may represent an acute exacerbation in those with established disease.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - Severe inflammation may result in mural ulcers, and penetrating complications such as inflammatory mesenteric mass, sinus tract, abscess, and fistula. The hallmark of Crohn’s disease is multifocal segmental involvement of the bowel with intervening areas of normal appearing bowel, also called ‘skip lesions’[9]. The terminal ileum is the most commonly involved part of the bowel followed by distal and mid ileum and ascending colon; however, any part of the small or large bowel can be affected. Long-standing Crohn’s disease can develop bowel strictures characterized by luminal narrowing of the involved segment and upstream bowel dilatation, sacculation, and intramural fat deposition. Mildly enlarged reactive mesenteric lymph nodes are commonly seen in Crohn’s; however, bulky lymphadenopathy is uncommon.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - Toxic megacolon may affect a focal colonic segment or involve the entire colon. Toxic megacolon is diagnosed as total or segmental non-obstructive colonic dilation (dilation of the colon more than 6 cm) with systemic toxic features. The systemic features required include at least 3 of the following: fever (> 38 degrees Celsius), pulse rate over 120beats/min, neutrophilic leucocytosis exceeding 10500/micro/l, or anemia. Furthermore, at least one of the following: dehydration, altered sensorium, electrolyte disturbances, and hypotension. Abdominal radiographs and computed tomography (CT) typically demonstrate marked colonic dilatation-defined as a transverse diameter exceeding 6 cm-often accompanied by air-fluid levels .
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - A plain abdominal radiograph may reveal a dilated colon with the classic thumbprinting sign, indicative of submucosal edema and inflammation . Once pseudomembranous colitis is clinically suspected, radiographs serve a valuable role in serial monitoring to assess for progressive colonic dilatation and to identify complications such as perforation or toxic megacolon. CT provides a more detailed assessment, typically demonstrating diffuse or segmental colonic dilatation with edematous mural thickening. Pan colitis, is usual with this particular bacterial infection. Thickened haustral folds may be visualized, producing the characteristic thumbprinting sign, and in some cases, the accordion sign, reflecting trapped oral contrast between thickened haustral folds. These features are characteristic of pseudomembranous colitis but not pathognomonic. Peri-colonic fat stranding and ascites are commonly present in ancillary features.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - Whipple disease is a multisystemic infection caused by the Tropheryma whipplei bacteria resulting in a constellation of insidious clinical features including abdominal pain, diarrhoea, weight loss, arthralgia, anemia, and lymphadenopathy. The classic imaging findings seen with Whipple disease are edematous wall thickening of the proximal small bowel, with or without nodular appearance; fluid distended bowel loops, characteristic low attenuation mesenteric lymphadenopathy and ascites. On ultrasound, the mesenteric lymphadenopathy appears typically hyperechoic, and the involved small bowel demonstrates loss of normal mural stratification. The gold standard test for diagnosis is histopathologic examination with periodic acid-Schiff (PAS) staining.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - GI tuberculosis accounts for 1 to 3% of all tuberculosis cases worldwide and is frequently caused by the dissemination of primary pulmonary infection; however, infrequently it can present as an isolated gastrointestinal infection.The ileocecal region is involved in 90% of the cases, however, any part of the gastrointestinal tract can be involved, including the colon and rectum. The classic CT features are circumferential asymmetric thickening of the ileocecal junction, terminal ileum and cecum that can lead to luminal narrowing and multifocal skip involvement of the small bowel similar to Crohn's disease. The presence of necrotic mesenteric necrotic lymphadenopathy, peritoneal thickening, peritoneal nodules, and ascites favour tuberculosis over Crohn’s disease. Peritoneal involvement can result in thickening of the peritoneum encasing the bowel loops, central clustering or matting of the bowel loops and obstruction
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - CT findings of parasite-infested intestine are non-specific with thickened, edematous bowel wall and effacement of mucosal folds. Worms such as Ascaris can be seen as elongated luminal filling defect in a contrast filled bowel on CT studies or as mobile echogenic linear intraluminal strips with acoustic shadowing on ultrasound. The worms typically demonstrate a "triple line" sign on ultrasound with a central anechoic tube (gut of the worm) between parallel echogenic lines. Parasitic infection of the GI tract can result in bowel obstruction, mesenteric inflammation and lymphadenopathy, and extra-intestinal complications like recurrent cholangitis, liver abscess, and pancreatitis due to the migration of worms into the biliary tract. Massive worm accumulation can result in occlusion of mesenteric vasculature, leading to intestinal infarction and gangrene.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - The classic finding of vasculitis includes abnormal wall thickening, fat stranding of the mesenteric arteries with or without vascular wall irregularity, luminal narrowing, and microaneurysms. On CT, the bowel findings of mesenteric vasculitis include circumferential thickening of the intestinal wall and mucosal folds resembling a ‘stack of coins’ , luminal dilation, submucosal edema, abnormal bowel enhancement, intramural hemorrhage, mesenteric edema, and signs of bowel ischemia in acute presentation . Vasculitis can result in lupus enteritis, which typically manifests as diffuse stratified thickening and mural enhancement of the small bowel due to submucosal edema, often accompanied by mesenteric edema and ascites.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - The incidence of immunotherapy-related enterocolitis is increasing, paralleling the growing use of immune checkpoint inhibitors in the treatment of various malignancies. Commonly implicated agents include Ipilimumab, Pembrolizumab, Nivolumab, Atezolizumab, Durvalumab, Avelumab, and Tremelimumab. Among these, colitis is more frequently observed than isolated small bowel enteritis, with diarrhea being the predominant clinical manifestation. On cross-sectional imaging, the most frequently encountered findings include segmental or diffuse colonic wall edema, luminal distension, mural hyperenhancement, and pericolic fat stranding—features reflective of active inflammation. Small bowel inflammation can present with diffuse inflammation or multifocal segmental inflammation mimicking Crohn’s disease. Optimal management of immune-mediated enterocolitis requires early recognition, and timely use of immunosuppressive agents, often requiring cessation of the immunotherapy in symptomatic patient.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - Arterial ischemia accounts for 40–50% cases of acute mesenteric ischemia and can occur due to arterial thrombosis or embolism. The most common underlying risk factors for emboli formation are cardiovascular conditions such as heart failure, myocardial infarction, cardiomyopathies, and atrial fibrillation [54, 55]. Superior mesenteric artery is the most susceptible artery for emboli, typically, distal to its first jejunal branch, due to acute angle origin from the aorta and high-velocity flow giving rise to a classic ischemic pattern that spares the proximal small intestine [56]. Risk factors for mesenteric ischemia due to arterial thrombosis include atherosclerosis, dyslipidemias, hypertension, diabetes and an estrogen-based oral contraceptive [55]
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - These account for nearly 10% of cases of intestinal ischemia. They can present up to 1–4 weeks after onset with non-specific symptoms such as abdominal pain and diarrhea. Nearly 60% of patients have a prior diagnosis of peripheral venous thrombosis or pulmonary embolism. Contrast-enhanced CT scan typically demonstrates a filling defect in the mesenteric vein with rim enhancement of the venous wall, often with prominent collateral veins in the mesentery. The bowel wall is significantly thickened in cases of venous ischemia, unlike the thinned-out wall seen in arterial-associated ischemia. The continuous inflow of arterial flow with an impeded venous outflow results in an increased intramural hydrostatic pressure and hypoattenuating edema with the submucosa layers presenting as the “target sign”. Due to the prolonged nature of the condition, there is usually prominent peri-enteric fat stranding and oedema.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - On routine portal-venous phase CT, signs of hypoperfusion complex include decreased enhancement of the solid organs such as spleen and liver with increased enhancement of the bowel mucosa due to splanchnic autoregulation, and hyperenhancement of the adrenal glands due to sympathetic response. In addition, there is reduction in the caliber of the aorta (less than 1.3 cm in diameter when measured approximately 1 cm below the origin of superior mesenteric artery), flattening of IVC (less than 9 mm AP diameter of the intrahepatic and infra-renal IVC) with circumferential zone of low attenuation fluid around the IVC and pancreas due to a hyper-permeable state secondary to a systemic inflammatory response syndrome, referred to as the “Halo sign”. Small bowel loops may be fluid-filled, occasionally dilated and thickened. The colon is typically normal appearing. At least 2 or more vascular, visceral, or parenchymal signs are necessary to establish the presence of a CT hypo perfusion complex.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press) - Features of neutropenic colitis seen on CT are circumferential wall thickening, peri-colonic fat stranding, and free-fluid predominantly in acute cases. Severe cases can show cecal pneumatosis, necrosis, mesenteric abscess formation, and perforation. Although it is difficult to differentiate it from other causes of enterocolitis, predominant inflammation of the right colon in a patient with neutropenia should raise concern for neutropenic colitis.
CT patterns of acute enterocolitis - a practical guide for the emergency radiologist
Snehal Rathi · Garima Suman · Avinash Nehra · Pranav Ajmera · Ashish Khandelwal
Emergency Radiology 2025 (in press)
Deep Learning
- Purpose: To evaluate the impact of a fully automated, multitask deep learning (DL) algorithm on interreader agreement of coronary artery disease (CAD) detection and stenosis classification using coronary CT angiography (CCTA).
Materials and Methods: This retrospective study included CCTA examinations (n = 623 patients) performed for clinical indications on CT systems from multiple vendors between January 2010 and December 2019. An expert reader (reader 1) analyzed all CCTA scans manually and with artificial intelligence (AI)–assisted reading at the lesion, coronary segment, and patient levels using the CAD Reporting and Data System (CAD-RADS). The AI algorithm detected, quantified, and classified coronary lesions. Interreader agreement was evaluated using a second expert reader (reader 2), who analyzed a randomly selected subset of 274 patients. CAD-RADS scores from radiologist reports (reader 3) were available for 362 patients. In a subgroup of 30 patients with disagreements, R2 also interpreted the cases using AI assistance. Agreement between readings, with and without AI, was assessed using Spearman correlation,and logistic regression and mixed models evaluated the impact of AI-assisted reading on CAD-RADS classification.
Conclusion: AI-assisted reading using a DL algorithm significantly improved interreader agreement for CAD-RADS classification at CCTA.
Impact of Deep Learning–based Artificial Intelligence
Assistance on Reader Agreement in Coronary CT Angiography Interpretation Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 - ■ Artificial intelligence (AI)–assisted reading of coronary CT angiography improved interreader agreement between two expert readers for Coronary Artery Disease Reporting and Data System (CADRADS) classification, with Spearman correlation coefficient increasing from 0.899 to 0.949 (P < .001).
■ In cases in which reader 1 and the AI disagreed, interreader agreement between reader 1 and reader 2 was low with manual readings (ρ = 0.688) but significantly improved when both readers used AI-assisted reading (ρ = 0.975; P < .001).
■ AI-assisted reading of coronary CT angiography improved the agreement between an expert reader and the existing radiologist reports from multiple radiologists for CAD-RADS classification, with the Spearman correlation coefficient increasing from 0.889 to 0.938 (P < .001).
Impact of Deep Learning–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 - Finally, the role of AI-assisted reading in CAD-RADS reporting is not limited to standardization and variability reduction. In our study, AI helped reader 1 identify 11 patients with at least one plaque (CAD-RADS ≥ 0) that were labeled as CAD-RADS 0 during manual reading. The detection of small, noncalcified, easy-to-miss plaques is important because the presence of a plaque, even a small one, can prompt clinicians to start preventive therapy . Moreover, detecting early signs of CAD at CCTA is important, especially for prognostic reasons. In the PROMISE (Prospective Multicenter Imaging Study for Evaluation of Chest Pain) trial, a large prospective trial of CCTA in patients with stable chest pain and suspected CAD, CAD-RADS showed a greater prognostic value than all other traditional scores tested. Because CCTA is progressively assuming the role of a gatekeeper for unnecessary invasive coronary angiography, one of the main purposes of AI introduction to cardiovascular imaging is to reduce the number of false negatives.
ng–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 - In conclusion, the use of a DL algorithm for AI-assisted reading significantly improved interreader agreement in CADRADS classification of CCTA images. Future studies should test AI-assisted reading on larger, more robust datasets with consistent and overlapping expert annotations. They should also assess its clinical impact, real-time performance across diverse settings, and potential role in training less experienced readers. Correlation with invasive coronary angiography may further support its diagnostic value.
Impact of Deep Learning–based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation
Roberto Farì, et al.
Radiology: Cardiothoracic Imaging 2025; 7(5):e240563 - AI and ML algorithms will continue to be developed for tasks such as nodule detection and diagnostic classification. Developing robust AI and ML models requires a sufficiently large training dataset that is representative of the intended use population (eg, balanced with respect to demographics such as sex, age, race, disease characteristics such as severity, and imaging equipment and modalities). Models may be further optimized using an independent validation set or by using cross-validation. To ensure the model is not overfit, performance of the developed model must be characterized using an independent test set that is also representative of the intended use population. In this study, the training dataset was so large that the Bayes limit was reached without using the entire sample, indicating that only improved independent features—not additional data—would be expected to further enhance model performance.
How Much Training Data Does AI Need?
Kellie J. Archer,
Radiology: Artificial Intelligence 2025; 7(6):e250685 - Data quality, good features, and representation are among the most important factors for AI and ML tasks . It is important for the reference standard to be reliable because label noise negatively impacts the apparent Bayes error rate. To this end, in this article, the authors annotated nodules in the NLST dataset for cancer cases. To accurately predict class labels, the measured features should have distributions that do not overlap by class. As seen in their Table S1, nodule size has very distinct distributions between malignant and benign nodules. The training data should span the full variability of the clinical application. The 16 077 nodules used for training were from 22 different scanners from four manufacturers, and participants were recruited from 33 sites. Although the training dataset did not match the racial distribution of the U.S. population, all major racial groups were represented, and there was good balance with respect to sex.
How Much Training Data Does AI Need?
Kellie J. Archer,
Radiology: Artificial Intelligence 2025; 7(6):e250685
- Health and health care AI tools should be subject to a governance structure that protects individuals and ensures the tools achieve their potential benefits. For other health care interventions, regulatory oversight is an important part of that governance, assuring society and markets that an intervention is credible. However, the US has no comprehensive fit-for-purpose regulatory framework for health and health care AI. Reasons include the diverse and rapidly evolving nature of AI technology, the numerous agencies with jurisdiction over different types and aspects of AI, and a lack of regulatory frameworks specifically tailored for AI within these agencies.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - The impact of AI on the health careworkforce will be wide-ranging. Clinicians may be excited by the potential benefits,worry about job displacement, and enjoy or resist requirements to improve their AI literacy. They may also have existential concerns like misalignment of human and AI ethos, goals, and principles. Although their comfort with AI is increasing, US physicians face a lack of regulatory oversight as the primary reason for their lack of trust and adoption of both clinical and health care business operations tools.Health care worker unions are also raising concerns about unsafe and underregulated AI. The workforce composition may also need to change, adding more experts in the development, implementation, and evaluation of AI tools. And, of course, AI tools have the potential to augment how many clinicians work, albeit with key caveats.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - Third, new technology is often first available to, and adopted by, individuals and organizations with greater means and resources. If AI tools are to be developed and disseminated in a manner that is fair, equitable, and does not widen the digital divide, then any education and reorganization efforts must include those parts of health care delivery responsible for the most vulnerable groups. Of course, efforts to ensure fair access to AI must also be cognizant of the risks of deploying a tool with potential untoward consequences in settings poorly equipped to detect them.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - Fourth, many AI tools are aimed at reducing the administrative burden on health care professionals (eg, medical record documentation or prior authorization appeals) on the premise that this burden contributes to burnout, low morale, and stress. However, this line of reasoning may have flaws. First, burnout, low morale, and stress are wicked problems; it is a tall ask to expect an AI tool to fix them. Second, if freed from administrative tasks, clinicians may be asked to see more patients, which could also cause burnout. Third, focusing on tools to automate tasks such as prior authorization potentially misses the larger opportunity to rethink entirely the purpose and value of such tasks
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - Though not discussed in detail here, important ethical and legal issues will affect adoption of health and health care AI. One issue is data rights, privacy, and ownership. Health and health care data are essential fuel for AI tools. Although one can ascribe where health and health care data originate, it is less clear who owns, or ought to own, the data, especially when aggregated and transformed, and what rights for privacy and use extend, or should extend, to whom. The Health Insurance Portability and Accountability Act intellectual property law provide some guidance on privacy rights, security obligations, trade secrets, and ownership of tools developed from data, but are less clear on ownership of underlying data and do not protect against emergence of dark markets, reflect all ethical considerations, or fully align with state or international legislation, such as the EU’s AI Act and General Data Protection Regulation.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - Finally, use of AI tools has thus far largely been voluntary. However, as their benefits become more established, failure to use an AI tool may be considered unethical or a breach of standard care.A health care system or professional may thus be liable in a malpractice suit for failing to use AI. At the same time, if a plaintiff sues for an adverse outcome when care was provided in which an AI tool was involved, the question arises of whether liability rests with the health care professional, the health care system,or the developer of the tool. Though relevant case law is currently limited, developers, health care systems, and health care professionals will all have to adopt strategies to manage their liability risk. These examples are just some of the many new issues that will need to be addressed as AI becomes more incorporated in health and health care.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - AI will massively disrupt health and health care delivery in the coming years. The traditional approaches to evaluate, regulate, and monitor novel health care interventions are being pushed to their limits, especially with generative and agentic AI, and especially because the tools’ effects cannot be fully understood until deployed in practice. Nonetheless, many tools are already being rapidly adopted, in part because they are addressing important pain points for end users. Given the many long-standing problems in health care, this disruption represents an incredible opportunity. However, the odds that this disruption will improve health for all will depend heavily on creation of an ecosystem capable of rapid, efficient, robust, and generalizable knowledge about the consequences of these tools on health.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - The past year marked a turning point for AI in clinical care, as ambient documentation tools and decision support algorithms moved decisively from pilot projects into widespread implementation across health care systems. Industry surveys indicate that approximately 70% of payers and providers in the United States are pursuing or have already implemented generative AI projects. This rapid adoption has begun reshaping clinical workflows, patient–clinician interactions, and reimbursement policies. Yet, as AI shifts from controlled research settings into complex real-world environments, both its benefits and its limitations have become more evident.
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
Pierre Elias, Kameron Collin Black , Payal Chandak et al.
NEJM Oct 2025 (in press) - Adoption of AI in clinical practice 2025
AI for Clinical Documentation
AI in Insurance Denials
AI for Clinical Decision Support
Clinician–AI Interaction
Clinician–AI Interaction
Clinician–AI Interaction - In a prospective German screening study (PRAIM study [Prospective Multicenter Observational Study of an Integrated AI System with Live Monitoring], 2025), radiologists using an AI-assisted double-reading workflow found 17.6% more breast cancers than standard double reads. The AI-guided group detected 6.7 cancers per 1000 screened versus 5.7 cancers per 1000 screened in controls, without raising false recalls. This indicates a synergistic effect: the computer complements, rather than replaces, human radiologists.
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
Pierre Elias, Kameron Collin Black , Payal Chandak et al.
NEJM Oct 2025 (in press) - In contrast, the value of clinician–AI collaboration in diagnostic reasoning remains less certain. Google DeepMind’s conversational multiagent system (Articulate Medical Intelligence Explorer, or AMIE) outperformed primary care physicians across diagnosis, management, communication, and empathy in standardized patient scenarios. Similarly, two recent randomized controlled trials showed Generative Pretrained Transformer 4 alone matched physicians using conventional or large language model (LLM)–supported resources in clinical management and surpassed them in diagnostic accuracy. Such results demonstrate AI’s potential to exceed clinician performance in knowledge-driven tasks, but there is a real need to ensure clinicians are calibrated to the value of new technologies and use them in real-world settings to determine how they truly fare.
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
Pierre Elias, Kameron Collin Black , Payal Chandak et al.
NEJM Oct 2025 (in press) - Although regulatory frameworks remain under debate, unregulated AI medical scribes continue to be deployed widely across health care systems. The U.S. Food and Drug Administration has yet to authorize an LLM, as it is unclear whether to evaluate them as new devices or “novel forms of intelligence” that undergo training and evaluation similar to physicians and nurses. Should LLMs be held to the same standards as expert human clinicians? Are humans the ideal examiners of LLMs’ behavior? Until a coherent regulatory pathway emerges, clinicians and health systems will shoulder the responsibility and liability for ensuring these tools benefit rather than harm patients.
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
Pierre Elias, Kameron Collin Black , Payal Chandak et al.
NEJM Oct 2025 (in press) - Clinical AI has passed an inflection point. AI has appeared in the back office and at the bedside, and it is here to stay. Every health system and clinician will likely have at least one AI technology they must reckon with in the coming year. Given this rapid proliferation, it is essential to evaluate the consequences of integrating AI into documentation and decision-making, lest we embed new risks rather than deliver lasting benefits. The field now faces three urgent tasks: first, to establish rigorous frameworks for evaluating AI safety, efficacy, and postdeployment performance; second, to conduct meaningful ROI assessments to guide adoption; and third, to redefine the roles of clinicians and institutions so that AI complements rather than undermines human judgment. Achieving these goals will require aligning market incentives with meaningful patient outcomes and developing transparent pathways for oversight. Whether we meet these requirements will determine if AI becomes a durable driver of better health or a short-lived experiment.
The Moment AI Arrived in the Clinic: Insights from the SAIL 2025 Year in Review
Pierre Elias, Kameron Collin Black , Payal Chandak et al.
NEJM Oct 2025 (in press) - Clinical AI has passed an inflection point. AI has appeared in the back office and at the bedside, and it is here to stay. Every health system and clinician will likely have at least one AI technology they must reckon with in the coming year. Given this rapid proliferation, it is essential to evaluate the consequences of integrating AI into documentation and decision-making, lest we embed new risks rather than deliver lasting benefits. The field now faces three urgent tasks: first, to establish rigorous frameworks for evaluating AI safety, efficacy, and postdeployment performance; second, to conduct meaningful ROI assessments to guide adoption; and third, to redefine the roles of clinicians and institutions so that AI complements rather than undermines human judgment. Achieving these goals will require aligning market incentives with meaningful patient outcomes and developing transparent pathways for oversight. Whether we meet these requirements will determine if AI becomes a durable driver of better health or a short-lived experiment.
- AI excels at pattern recognition. As large language models and their training data improve, AI will likely outperform human diagnosticians, especially in routine cases that cohere with prototypical illness scripts. Human physicians should develop complementary skills that machines can less easily replicate, such as flexible reasoning, creative problem-solving, and the ability to navigate uncertainty in cases involving new knowledge and/or unfamiliar presentations. To the extent that medicine represents a “wicked” environment—marked by complexity, incomplete rules, and delayed or inaccurate feedback—it is less amenable to AI, suggesting that efforts to teach clinical reasoning should prioritize flexibility and nuance over automaticity.
Critical Thinking for 21st-Century Medicine—Moving Beyond Illness Scripts
Richard M. Schwartzstein, MD; Alexander A. Iyer, ScB
JAMA November 4, 2025 Volume 334, Number 17 - For decades, educators have attempted to reduce medical errors by improving learners’ clinical reasoning skills.The teaching of illness scripts has traditionally been central to these efforts. An illness script is a clinician’s organized mental summary of their knowledge about a disease, encompassing key elements including risk factors, pathophysiology, and clinical consequences. It functions as a mental flash card that helps a clinician rapidly recall and apply knowledge to diagnose patients. Illness scripts rely on pattern recognition and are argued to reflect the thought processes of experts, offering an efficient framework for teaching learners to create differential diagnoses. However,do illness scripts actually improve the types of thinking that allow expert clinicians to reduce diagnostic error and improve patient outcomes? Is there a better model for clinical reasoning?
Critical Thinking for 21st-Century Medicine—Moving Beyond Illness Scripts
Richard M. Schwartzstein, MD; Alexander A. Iyer, ScB
JAMA November 4, 2025 Volume 334, Number 17 - Approaches to clinical reasoning based heavily on illness scripts can be thought of as cultivating routine expertise. Routine experts have large repositories of illness scripts on which to draw when evaluating patients, making them efficient and reliable in many clinical scenarios. However, when confronted with a new problem, routine experts attempt to adapt the problem to solutions with which they are comfortable rather than flexibly creating solutions that better match the problem. This means that evenwhena patient’s presentation does not fully match an illness script, a routine expert willdefault to pattern-oriented thought processes, increasing the likelihood of a diagnostic error.
Critical Thinking for 21st-Century Medicine—Moving Beyond Illness Scripts
Richard M. Schwartzstein, MD; Alexander A. Iyer, ScB
JAMA November 4, 2025 Volume 334, Number 17 - Now more than ever physicians must build their practices on the foundation of strong critical thinking skills. In the 21st century, this means medical education must go beyond teaching students what to think and instead teach students how to think when patterns do not fit (or how to verify when patterns do seem to fit). By questioning a primary focus on illness scripts, emphasizing pathophysiological reasoning, and cultivating adaptive expertise, we can prepare future physicians—and their patients—for what may lie ahead.
Critical Thinking for 21st-Century Medicine—Moving Beyond Illness Scripts
Richard M. Schwartzstein, MD; Alexander A. Iyer, ScB
JAMA November 4, 2025 Volume 334, Number 17 - Radiomics uses machine learning to capture detailed interpixel or intervoxel relationships within medical images, beyond the limits of human perception and current radiology reporting. This potential to more deeply explore the content of medical images has fueled a growing interest in radiomics by the radiology research community, as investigators have sought to use radiomics to provide novel insights and biomarkers to improve patient outcomes. However, further integration of radiomics into clinical practice has been hindered by a range of challenges and barriers, precluding full clinical adoption of any single radiomics application. This article describes current limitations in the existing body of radiomics research and highlightsthe strategic changes needed for radiomics research to yield more meaningful translational evidence in this space.
- IMPORTANCE Artificial intelligence (AI) is changing health and health care on an unprecedented scale. Though the potential benefits are massive, so are the risks. The JAMA Summit on AI discussed how health and health care AI should be developed, evaluated, regulated, disseminated, and monitored.
CONCLUSIONS AND RELEVANCE AI will disrupt every part of health and health care delivery in the coming years. Given the many long-standing problems in health care, this disruption represents an incredible opportunity. However, the odds that this disruption will improve health for all will depend heavily on the creation of an ecosystem capable of rapid, efficient, robust, and generalizable knowledge about the consequences of these tools on health.
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025. - The following 3 advances help differentiate AI from prior digital technologies:
• Deep learning: development of deeper, more convoluted neural networks capable of interpreting large complex datasets to address specific yet complicated tasks (eg, computer vision).
• Generative AI: an extension of deep learning using so-called large language and foundation models capable of generating new content to address far broader task requests (eg, ChatGPT or Gemini).
• Agentic AI: an extension of deep learning and generative AI capable of autonomous decision-making (eg, the Tesla autopilot software for autonomous driving).
AI, Health, and Health Care Today and Tomorrow
The JAMA Summit Report on Artificial Intelligence
Derek C. Angus, MD, MPH; Rohan Khera, MD, MS; Tracy Lieu et al.
JAMA. doi:10.1001/jama.2025.18490 Published online October 13, 2025.
- Radiomics uses machine learning to capture detailed interpixel or intervoxel relationships within medical images, beyond the limits of human perception and current radiology reporting. This potential to more deeply explore the content of medical images has fueled a growing interest in radiomics by the radiology research community, as investigators have sought to use radiomics to provide novel insights and biomarkers to improve patient outcomes. However, further integration of radiomics into clinical practice has been hindered by a range of challenges and barriers, precluding full clinical adoption of any single radiomics application. This article describes current limitations in the existing body of radiomics research and highlights the strategic changes needed for radiomics research to yield more meaningful translational evidence in this space. Though setting a high bar, the guidance could lead to the evidence required for radiomics to move into clinical arenas—or indicate if radiomics best stays a research technique.
Radiomics Research: Current Status, Limitations, and Guide for Strategic Changes
Hyun Soo Ko, Andrew B. Rosenkrantz
Roentgen Ray Rev 2025; 1:e2501096ISSN 3068-0301/25/14–e2501096 - Radiomics uses machine learning to analyze medical images and identify patterns beyond that which a human could perceive through a complex process of image acquisition and processing, lesion segmentation, feature extraction, and model development and testing. These patterns are of clinical interest given a presumed opportunity to elicit unique information that informs clinical decision-making. Such information is based on radiomic features (RFs) that capture interpixel and intervoxel relationships in more detail and complexity than do the simple metrics typically found in radiology reports (e.g., lesion size or attenuation). This potential clinical relevance of radiomics—to more deeply explore the content of medical images for purposes such as characterizing lesions, predicting behavior, and detecting mutations that drive growth—has stimulated exponential growth in radiomics research over the past 2–3 decades, as investigators have sought to extract novel RFs from noninvasive examinations for guiding diagnosis, risk stratification, and treatment selection.
Radiomics Research: Current Status, Limitations, and Guide for Strategic Changes
Hyun Soo Ko, Andrew B. Rosenkrantz
Roentgen Ray Rev 2025; 1:e2501096ISSN 3068-0301/25/14–e2501096 - A well-recognized barrier to the generalizability of existing radiomics models is a lack of diversity commonly encountered in the training datasets used for model development, typically originating from one institution or a small number of institutions. Such datasets have limited heterogeneity with respect to both patient demographic characteristics (e.g., age, sex, ethnicity, geography, environment) and disease characteristics (e.g., the spectrum of pathologic entities represented, as well as the stage, severity, type, or other features of the presented conditions). Another source of nonuniformity of research datasets relates to the equipment (both hardware and software) used. For example, the dataset is unlikely to represent the full spectrum of scanner vendors and models.
Radiomics Research: Current Status, Limitations, and Guide for Strategic Changes
Hyun Soo Ko, Andrew B. Rosenkrantz
Roentgen Ray Rev 2025; 1:e2501096ISSN 3068-0301/25/14–e2501096 - Moreover, because radiomics models are highly fit to the datasets on which they learn, the results may not successfully translate when applied to other patient populations having distinct demographic and disease characteristics, evaluated using varying equipment and protocols, and with interpretations by different readers. A further issue hindering generalizability relates to inconsistent results depending on the software used for radiomics analysis and RF generation. A wide array of commercial and publicly available radiomics analysis platforms have arisen, yielding distinct values for a given RF when processing the same image set.
Radiomics Research: Current Status, Limitations, and Guide for Strategic Changes
Hyun Soo Ko, Andrew B. Rosenkrantz
Roentgen Ray Rev 2025; 1:e2501096ISSN 3068-0301/25/14–e2501096 - Several additional steps could improve the potential clinical utility of radiomics models. First, fully automated methods to extract RFs would avoid the need for human annotation. Second, integration of radiomics analysis with advanced artificial intelligence methods, such as supervised deep learning, could create pipelines that are more flexible and less dependent on idiosyncratic characteristics of the training data. Third, explanations of how models produce their outputs would provide radiologist users with a greater understanding and ability to more effectively communicate feedback on model performance. Guidelines and checklists, such as the Checklist for Evaluation of Radiomics Research (CLEAR) , the Methodological Radiomics Score (METRICS), and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM), aim to serve as educational tools for researchers and reviewers to improve transparency, reproducibility, and generalizability of radiomics research. Such understanding is crucial for models to become embedded into clinical decision-making. Finally, applying models in environments that allow ongoing postdeployment evaluation would allow the model’s continuous learning and adaptation once exposed to real-world clinical environments.
Radiomics Research: Current Status, Limitations, and Guide for Strategic Changes
Hyun Soo Ko, Andrew B. Rosenkrantz
Roentgen Ray Rev 2025; 1:e2501096ISSN 3068-0301/25/14–e2501096 - In summary, no current radiomics model is ready for widespread clinical adoption, as research strives to improve workflow consistencies and eliminate sources of bias, including demographic, geographic, and technologic differences among health care settings. Although setting a high bar, a prospective multicenter study with evaluation across numerous diverse external test sets, incorporating automated RF extraction and supervised deep learning, could provide the evidence needed to truly advance radiomics into the clinical arena—or indicate if radiomics best remains a research tool.
Radiomics Research: Current Status, Limitations, and Guide for Strategic Changes
Hyun Soo Ko, Andrew B. Rosenkrantz
Roentgen Ray Rev 2025; 1:e2501096ISSN 3068-0301/25/14–e2501096 - The risk is that the approach becomes too rapid, embedding AI and making it autonomous in high-stakes decision-making, as in the case of agentic AI, before supporting, high-certainty evidence is mature. This could not only cause harm, but also weaken the trust that underpins sustained adoption. Conversely, in maintaining too great a distance, even from well-validated applications, we risk missing opportunities to improve patient care and relieve clinician burden, to optimize daily clinical reasoning, and guide diagnostic and therapeutic decision-making at the bedside. Clinical validation is essential to avoid misjudging how closely and how fast medicine should approach AI.
Letter: The Hedgehog’s Dilemma and the Integration of Artificial Intelligence in Medicine
Alberto Rizzo , Daniele Roberto Giacobbe , and Alberto Dolci
NEJM AI 2025 DOI: 10.1056/AIp2500952 - Recent consensus recommendations highlight traceability, robustness, and governance, among others, as essential elements for trustworthy deployment. These align with Taylor’s call for value alignment, linking technical validation with safeguards that matter to clinicians and patients.Nevertheless, evidence and governance are not enough on their own. Clinicians and researchers also need training to work with AI systems, supported by procedures for ongoing monitoring and adaptation. Such measures recognize that the balance between closeness and distance is not static, but must evolve with new data and experience. The hedgehog’s dilemma is a reminder that integration of AI into medicine requires careful calibration rather than either rapid convergence or persistent distance. A gradual, risk-based approach can support innovation while preserving trust and safety.
The Hedgehog’s Dilemma and the Integration of Artificial Intelligence in Medicine
Alberto Rizzo , Daniele Roberto Giacobbe , and Alberto Dolci
NEJM AI 2025 DOI: 10.1056/AIp2500952 - We share their caution that untested autonomy in high-stakes settings can erode trust. Yet, in health systems, high-certainty evidence often accrues locally through learning-health-system cycles (e.g., site-specific antibiotic stewardship). Accordingly, risk should match local capability — begin in advisory-only mode with close oversight, measure value realization continuously, and escalate or de-escalate privileges as evidence accrues. Concretely, agent actions should be gated by prespecified thresholds and guardrails across the value portfolio (safety, effectiveness, equity, experience, and cost), with automatic rollback to human-only control when drift or harm signals emerge.
Response to “The Hedgehog’s dilemma and the Integration of Artificial Intelligence in Medicine”
R. Andrew Taylor
NEJM AI 2025 DOI: 10.1056/AIp2501053 - Rather than mandating default human interpretability, we favor transparency of ends and evidence. Agentic systems should make explicit the values they optimize, the metrics that instantiate those values, and the extent to which outcomes conform. Because the link from strategic aims to technical predicates is fragile, we need value traceability — a verifiable chain from declared priorities to measurements to actions. Developing such computational governance — not enforcing glass-box algorithms — is the pressing problem.
Response to “The Hedgehog’s dilemma and the Integration of Artificial Intelligence in Medicine”
R. Andrew Taylor , M.D., M.H.S.1
NEJM AI 2025 DOI: 10.1056/AIp2501053
- Advancements in artificial intelligence (AI) are transforming medical imaging diagnostics, offering new possibilities for automated pancreatic tumor detection in computed tomography scans. Pancreatic ductal adenocarcinoma continues to be one of the most lethal malignancies, with early detection being critical for improving survival rates. Deep learning models can learn hierarchical feature representations directly from imaging data, enhancing tumor detection accuracy. However, variations in model performance, impaired generalizability, and limited interpretability remain critical barriers to clinical adoption. This article provides a comprehensive overview of deep learning-based pancreatic tumor detection, discussing fundamental concepts, recent advancements, and challenges for clinical adoption. Implementation of deep learning tumor detection models into imaging workflows holds promise for improving early detection rates of pancreatic tumors. Addressing issues of standardization, external validation, and model transparency will be essential to enable the integration of AI into pancreatic cancer screening and diagnostics, ultimately improving early detection and patient outcomes.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - In the early phases of AI-driven tumor detection, 2 major approaches took the lead in the quantitative analysis of medical imaging features: radiomics-based machine learning models and deep learning (DL) approaches. Both techniques have shown promise in the detection of different pancreatic tumors While both extract quantitative features from medical images, they do so in fundamentally different ways. Radiomics models rely on predefined, handcrafted features derived from imaging data—such as shape, texture, and intensity-based metrics—requiring expert-driven feature selection and statistical analysis. In contrast, DL models automatically learn hierarchical features directly from the images, capturing complex patterns and learning connections without the need for manual feature engineering.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Classification is a supervised learning task, meaning the user has control over the input data and the corresponding expected output during training. In this context, the goal is to train an algorithm to recognize patterns in labeled data to categorize the input (eg, CT scans) into predefined categories (eg, “normal pancreas” vs “PDAC”). DL classification models extract relevant features using operations such as convolutional layers, which learn to identify distinguishing patterns from labeled training examples. Convolutional neural networks (CNNs) leverage these learning mechanisms to recognize subtle variations in texture and density that may not be immediately visible to the human eye, and in the case of tumor detection, might indicate malignancies.20Once trained, these models can generalize learned features to new cases, enabling classification of unseen images.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - A reliable assessment of classification models requires splitting the data into training/validation sets for development, and reserving a separate testing set to evaluate generalization. This can determine the presence of overfitting, where models perform well on training data but fail on new unseen cases. Additionally, performing an external validation using independent datasets provides a more robust evaluation of performance and generalizability across diverse populations or acquisition conditions. Classification models for tumor detection are often assessed using several complementary key metrics. Accuracy measures the overall correctness of the classification (ie, distinguishing tumor vs no tumor cases); however, this measure alone can be misleading in highly imbalanced tasks, such as detecting low-incidence tumors. Sensitivity, or recall, evaluates the ability of a model to detect cases with tumors (true positive rate), while specificity quantifies its ability to correctly identify negative controls (true negative rate).
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) 
- While any form of tumor detection can be valuable, models that perform well only on large, hard-to-miss tumors offer minimal benefit to the diagnostic workflow. Despite advancements in automated tumor diagnosis using DL models, detecting small pancreatic tumors in medical imaging remains one of the greatest challenges. Small lesions often exhibit only subtle visual cues that can be easily overlooked, even by experienced radiologists.Although the definition of a small tumor can be somewhat arbitrary, it is most commonly defined as one with the largest diameter of less than 2 centimeters (ie T1 stage). This threshold has demonstrated prognostic significance, with higher chances of tumor resectability and lower risk of metastasis at presentation.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Despite significant advancements in DL models for pancreatic tumor detection, several challenges remain before widespread clinical implementation can be achieved. A major challenge models face in achieving clinical applicability is generalizing performance beyond the training dataset. For instance, Liu et al. observed a significant performance decline in their CNN classification model, trained exclusively on an Asian population, when evaluated on an external cohort from the United States. They attributed the performance drop to differences in population characteristics and scanning parameters, highlighting the variability in model robustness across diverse patient groups.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - To address these challenges, standardization initiatives such as CLAIM (Checklist for Artificial Intelligence in Medical Imaging) aim to establish guidelines for model development, validation, and reporting, ensuring greater consistency across AI models.47,48By harmonizing evaluation protocols and promoting transparent reporting, these initiatives help bridge the gap between research and clinical implementation of DL-based tumor detection models. Other efforts, such as openly sharing code and final model versions alongside published results, further enhance model evaluation, reproducibility, and continuous model improvement.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - To address this challenge, ongoing efforts in Explainable AI focus on techniques such as saliency maps, feature attribution and attention mechanisms, which help visualize the most influential regions within an image that drive the prediction.50By enhancing transparency, these methods improve interpretability and foster clinician trust in DL-based diagnostic models.51Integrating explainability frameworks, improving cross-cohort model validation, and aligning AI models with standardized reporting guidelines will be essential to ensure reproducibility, robustness, and move toward the widespread clinical adoption of automated pancreatic tumor screening.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - One promising direction in multimodal detection is the analysis of longitudinal electronic health records (EHR). Emerging applications of LLMs show strong potential to extract information from unstructured clinical narratives and temporal trends, generating representations that can be interpreted by AI systems.These models can potentially identify subtle and evolving signals, such as symptom patterns, laboratory value changes, or medication adjustments, that may appear before anatomical abnormalities become detectable. For instance, in the setting of pancreatic cancer, the onset of new or worsening diabetes has been recognized as a potential early marker of disease.AI models may be able to detect concerning patterns in advance and integrate them with imaging and molecular data into multimodal prediction frameworks, enabling earlier risk stratification and more effective diagnostic and monitoring strategies.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Another promising area of research is the integration of CT detection models with liquid biopsy techniques. Liquid biopsy provides a minimally invasive approach to detect tumor-derived material in blood or other body fluids, including circulating tumor DNA (ctDNA), circulating tumor cells, and exosomal markers. Although more commonly used for disease monitoring, these biomarkers have also demonstrated potential for identifying early molecular changes related to pancreatic cancer, often before abnormalities can be seen on imaging studies.4Among these, ctDNA has garnered significant attention due to its ability to detect clonal somatic mutations, such as the KRAS mutations commonly found in pancreatic tumors.Studies have shown that combining ctDNA analysis with protein biomarkers (eg, CA19-9, CEA, HGF, and osteopontin) can significantly improve diagnostic sensitivity for pancreatic cancer.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - The future of early pancreatic cancer detection will depend on developing multimodal AI systems that are trained on rich and diverse sources of information, encompassing imaging, clinical records, molecular profiling, and systemic biomarkers. As these technologies evolve, rigorous validation, adherence to regulatory standards, and seamless integration into existing radiology workflows will be necessary to build trust among physicians and fully realize the potential of AI-driven pancreatic tumor detection in clinical practice.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Question What are the associations of CT-evaluated body composition with early recurrence and overall survival in patients with pancreatic ductal adenocarcinoma after resection?
Findings High visceral-to-subcutaneous fat ratio is an independent predictor for early recurrence, whereas high skeletal muscle density and subcutaneous fat area independently predict better overall survival.
Prognostic value of body composition on early recurrence and long-term survival of resectable pancreatic ductal adenocarcinoma
Linxia Wu, Tong Nie, Xiaoling Zhi et al.
European Radiology 2025 (in press) https://doi.org/10.1007/s00330-025-12028-8 - Results High VSR was an independent predictor for ER (OR: 2.304, p=0.001) and worse OS (HR: 1.462, p=0.007), whereas high SMD (HR: 0.609, p=0.005) and high SFA (HR: 0.649, p=0.002) were independent predictors for better OS. Subgroup analyses revealed variations in the prognostic effect of VSR according to diabetes status and tumor size. A model combining body composition metrics and clinicopathological indicators (carbohydrate antigen 19-9, carbohydrate antigen 12-5, tumor-node-metastasis stage, lymphovascular invasion, and adjuvant therapy) demonstrated good predictive ability for ER, with AUCs of 0.80 in the training set and 0.82 in the validation set.
Conclusion High VSR was an independent predictor for ER and worse OS in PDAC. Moreover, combining bod composition metrics and clinicopathological indicators can improvetheprognosis prediction of patients with PDAC after surgery.
Prognostic value of body composition on early recurrence and long-term survival of resectable pancreatic ductal adenocarcinoma
Linxia Wu, Tong Nie, Xiaoling Zhi et al.
European Radiology 2025 (in press) https://doi.org/10.1007/s00330-025-12028-8 - This study highlighted the utility and prognostic value of preoperative CT-based body composition in predicting ER and overall survival in patients with PDAC. The integrated model combining body composition parameters with clinicopathological indicators, including CA 19-9, CA 12-5, TNM stage, lymphovascular invasion, and adjuvant therapy, demonstrated good predictive performance for patient prognosis. These findings could help identify patients at a high risk of ER and assist clinicians in delivering personalized management and trpeatment strategies.
Prognostic value of body composition on early recurrence and long-term survival of resectable pancreatic ductal adenocarcinoma
Linxia Wu, Tong Nie, Xiaoling Zhi et al.
European Radiology 2025 (in press) https://doi.org/10.1007/s00330-025-12028-8 - Our findings show that AI trained on large, diverse datasets can exceed average radiologist performance in PDAC detection on routine contrast-enhanced CT, providing a foundation for regulatory dialogue and prospective validation. We further show the generalisability of these findings by validating AI performance on two independent external cohorts from different countries that were not used for model training. Future validation trials are required to further evaluate AI performance across broader clinical settings and patient populations, as well as to assess workflow integration and assisted reading scenarios. Ongoing in-depth analyses of AI and reader performance stratified by reader experience, tumour size, and clinical stage will be the subject of a forthcoming study aimed at clarifying how AI can most effectively support clinicians in practice
- Background Pancreatic ductal adenocarcinoma (PDAC) has the worst prognosis among major cancer types, primarily due to late diagnosis on contrast-enhanced CT. Artificial intelligence (AI) can improve diagnostic performance, but robust benchmarks and reliable comparison to radiologists’ performance are scarce. We established an open-source benchmark with the aim of investigating AI systems for PDAC detection on CT and compared them to radiologists’ performance, at scale.
Interpretation AI demonstrated substantially improved PDAC detection on routine CT scans compared to radiologists on average, showing potential to detect cancer earlier and improve patient outcomes.
Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study
Natalia Alves et al.
Lancet Oncology 2025 (in press) - Purpose This study investigates the frequency, progression, and clinical implications of pancreatic findings on chest low-dose computed tomography (LDCT) scans performed for lung cancer screening.
Conclusion Incidental pancreatic findings were uncommon (0.9%) and included calcifications, atrophy/fatty infiltration, cysts, ductal dilatation, and masses. These findings do not by themselves indicate pancreatic cancer but warrant documentation and, when suspicious, dedicated pancreatic imaging. Radiologist scrutiny could improve detection accuracy, indicating the potential of a LDCT lung cancer screening program for detecting and monitoring pancreatic lesions.
Pancreatic findings in participants in a program of low-dose computed tomography screening for lung cancer
Gros, Louisa,b; Yip, Rowenaa; Zhu, Yeqinga; Li, Pengfeia; Paksashvili, Natelaa; Sun, Qia; Yankelevitz, David F.a; Henschke, Claudia I.a
European Journal of Cancer Prevention ():10.1097/CEJ.0000000000000997, November 18, 2025 - Out of 9467 participants, 90 (0.9%) had pancreatic findings, mostly male (54.4%), median age 64.7, with smoking (92.2%), alcohol use (41.1%), and diabetes (22%). Of these, 60 (66.7%) were detected on baseline LDCT, primarily as calcifications (73.3%), atrophy/fatty infiltration (18.3%), and duct dilatation (5%). Of the 90 participants, 27 underwent only baseline LDCT. Among the remaining 63, 33 had pancreatic findings on baseline scans, 27 of whom (81.8%) showed consistent findings on follow-up, and 30 developed pancreatic findings during surveillance. Rereview of the baseline scans showed that 68 participants (75.6%) had findings, including eight missed earlier. More cases of atrophy/fatty infiltration and other findings were detected compared to the original report, with calcifications remaining predominant (50 participants). Similar patterns were observed during the rereview of the latest LDCT scans. Two participants with detected lesions underwent biopsy, diagnosing a serous cystadenoma and pancreatic adenocarcinoma. The latter succumbed to pancreatic cancer.
Pancreatic findings in participants in a program of low-dose computed tomography screening for lung cancer
Gros, Louisa,b; Yip, Rowenaa; Zhu, Yeqinga; Li, Pengfeia; Paksashvili, Natelaa; Sun, Qia; Yankelevitz, David F.a; Henschke, Claudia I.a
European Journal of Cancer Prevention ():10.1097/CEJ.0000000000000997, November 18, 2025 - In 2023, a deep learning algorithm was trained on over 3000 noncontrast CT scans to identify PCs and seven subtypes of non- PC lesions, with an AUC over 0.98 for lesion detection when applied to a multicenter validation dataset of over 6,000 patients. The algorithm significantly outperformed human readers, even when readers were provided contrast-enhanced CT scans. When the algorithm diagnosis probabilities were used as an aid to human readers, their performance identifying PCs was significantly improved. The authors additionally showed an ability to identify PCs on noncontrast chest CTs. With further real-world validation, such tools may be a way to utilize common imaging studies obtained for non-PC reasons to inform the selection of individuals for further assessment.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - Over the past decade, there has been substantial interest in using ML and AI to enhance the diagnostic utility of pancreatic imaging studies. A 2022 study evaluated four ML classifiers trained on a dataset of volumetric pancreas segmentations from patients with PC with prediagnostic contrast-enhanced CT scans along with matched controls. The best-performing of these models (support vector machine) correctly classified 95% of prediagnostic and 90% of control CT scans in the test set and showed a higher discrimination performance than did two radiologist readers.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - The DNA- and RNA-based PancreaSeq test, developed by then University of Pittsburgh (Pittsburgh, Pennsylvania, USA), initially evaluated 74 PC-related genetic alterations to identify high-risk cysts. Next, the test was narrowed to assess alterations in gene or mRNA expression of KRAS, GNAS, BRAF, TP53, PRKACA/B, ALK, NTRK1/3, RET, SMAD4, and CEACAM5 to determine whether a cyst was neoplastic. PancreaSeq showed 82% sensitivity and100% specificity in detecting advanced HGD and early cancer in pancreatic cysts. Of note, the test provides information on cyst type and the risk of progression to high-grade dysplasia or cancer and is currently implemented in select institutes.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - Furthermore, screening regimens should not be considered independent of cost, especially in settings limited by patient resources and/or availability of equipment and expertise, therefore, the development of cost-effective approaches is important. To advance the field of PC early detection, large international groups such as the CAPS research program and the PRECEDE Consortium will be necessary to enroll enough patients to adequately power PC early detection studies. The efforts of these groups, along with the many researchers working to better understand the biology of PC, may drive a shift toward improved outcomes for this challenging disease.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942
Kidney
- Renal Cell Carcinoma 2025: American Cancer Society
About 80,980 new cases of kidney cancer (52,410 in men and 28,570 in women) will be diagnosed.
About 14,510 people (9,550 men and 4,960 women) will die from this disease
Most people with kidney cancer are older. The average age of people when they are diagnosed is 65, with most people being diagnosed between ages 55 and 74. Kidney cancer is uncommon in people younger than age 45. - Renal Cell Carcinoma 2025
- Renal Cell Carcinoma 2025
OB GYN
- Ovarian cancer remains the sixth most common cause of cancer mortality in women in the United States and is a leading cause of mortality among patients with gynecologic malignancies. Imaging plays an important role in pretreatment staging of epithelial ovarian cancers, the evaluation of posttreatment response, and follow-up. Accurate pretreatment imaging is integral to determine appropriate first-line therapy. By delineating the extent of disease, imaging can assist decision making regarding the likelihood of optimal primary cytoreduction or need for neoadjuvant chemotherapy when optimal cytoreduction is not felt to be achievable. Contrast-enhanced CT serves as a mainstay modality for the pretreatment assessment of ovarian cancer, with MRI, PET/CT, and, in some instances, PET/ MRI used in the pretreatment setting. CT and PET/CT are also integral to assessing response, including in the suspected recurrence setting, with MRI and PET/MRI being used in select cases.
ACR Appropriateness Criteria® Staging and Follow-Up of Ovarian Cancer: 2025 Update
Expert Panel on GYN and OB Imaging: Erica B. Stein,et al.
J Am Coll Radiol 2025;22:S689-S698 
- There are several different histopathologic subtypes of ovarian cancer, with epithelial ovarian cancer being the most common and accounting for approximately 90% of all malignant ovarian neoplasms. Most of the information in this document regarding imaging use for staging and recurrence applies to epithelial ovarian cancers. There are two major subtypes of epithelial ovarian cancers that are distinguished by molecular, genetic, and morphologic characteristics. Type I is the more indolent form that includes low-grade serous, low-grade endometrioid, mucinous tumors, and clear cell carcinomas. Type II includes aggressive neoplasms, such as- high-grade serous or endometrioid, and undifferentiated cancer. The aggressive, or type II ovarian cancers, typically present in advanced stages (stage III-IV), after the disease has spread beyond the pelvis.
ACR Appropriateness Criteria® Staging and Follow-Up of Ovarian Cancer: 2025 Update
Expert Panel on GYN and OB Imaging: Erica B. Stein,et al.
J Am Coll Radiol 2025;22:S689-S698 - Contrast enhanced CT is the most useful procedure in the preoperative evaluation of ovarian cancer. It can provide clinically relevant information, including assessment of locoregional tumor extent and distant sites of disease including peritoneum, omentum, mesentery, liver, and lymph nodes. Contrast-enhanced CT has a reported accuracy for ovarian cancer staging of up to 94%, and accurate abdominopelvic disease assessment can predict successful surgical cytoreduction. The sensitivity of CT staging varies depending on the anatomical location being examined. One of the significant drawbacks of CT in staging ovarian cancer is its limited ability to consistently identify tumor implants on the bowel surface, mesentery, or peritoneum that are smaller than 5 mm, particularly in the absence of ascites .
ACR Appropriateness Criteria® Staging and Follow-Up of Ovarian Cancer: 2025 Update
Expert Panel on GYN and OB Imaging: Erica B. Stein,et al.
J Am Coll Radiol 2025;22:S689-S698 - For initial pretreatment staging of ovarian cancer, CT abdomen and pelvis with IV contrast and CT chest with IV contrast are recommended complementary studies to stage the tumor and evaluate for distant metastases. MRI abdomen and pelvis without and with IV contrast, MRI abdomen and pelvis without IV contrast, CT abdomen and pelvis without IV contrast, CT chest without IV contrast, FDG-PET/MRI, and FDG-PET/CT may be appropriate in staging ovarian cancer before treatment.
ACR Appropriateness Criteria® Staging and Follow-Up of Ovarian Cancer: 2025 Update
Expert Panel on GYN and OB Imaging: Erica B. Stein,et al.
J Am Coll Radiol 2025;22:S689-S698
Pancreas
- Advancements in artificial intelligence (AI) are transforming medical imaging diagnostics, offering new possibilities for automated pancreatic tumor detection in computed tomography scans. Pancreatic ductal adenocarcinoma continues to be one of the most lethal malignancies, with early detection being critical for improving survival rates. Deep learning models can learn hierarchical feature representations directly from imaging data, enhancing tumor detection accuracy. However, variations in model performance, impaired generalizability, and limited interpretability remain critical barriers to clinical adoption. This article provides a comprehensive overview of deep learning-based pancreatic tumor detection, discussing fundamental concepts, recent advancements, and challenges for clinical adoption. Implementation of deep learning tumor detection models into imaging workflows holds promise for improving early detection rates of pancreatic tumors. Addressing issues of standardization, external validation, and model transparency will be essential to enable the integration of AI into pancreatic cancer screening and diagnostics, ultimately improving early detection and patient outcomes.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - In the early phases of AI-driven tumor detection, 2 major approaches took the lead in the quantitative analysis of medical imaging features: radiomics-based machine learning models and deep learning (DL) approaches. Both techniques have shown promise in the detection of different pancreatic tumors While both extract quantitative features from medical images, they do so in fundamentally different ways. Radiomics models rely on predefined, handcrafted features derived from imaging data—such as shape, texture, and intensity-based metrics—requiring expert-driven feature selection and statistical analysis. In contrast, DL models automatically learn hierarchical features directly from the images, capturing complex patterns and learning connections without the need for manual feature engineering.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Classification is a supervised learning task, meaning the user has control over the input data and the corresponding expected output during training. In this context, the goal is to train an algorithm to recognize patterns in labeled data to categorize the input (eg, CT scans) into predefined categories (eg, “normal pancreas” vs “PDAC”). DL classification models extract relevant features using operations such as convolutional layers, which learn to identify distinguishing patterns from labeled training examples. Convolutional neural networks (CNNs) leverage these learning mechanisms to recognize subtle variations in texture and density that may not be immediately visible to the human eye, and in the case of tumor detection, might indicate malignancies.20Once trained, these models can generalize learned features to new cases, enabling classification of unseen images.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - A reliable assessment of classification models requires splitting the data into training/validation sets for development, and reserving a separate testing set to evaluate generalization. This can determine the presence of overfitting, where models perform well on training data but fail on new unseen cases. Additionally, performing an external validation using independent datasets provides a more robust evaluation of performance and generalizability across diverse populations or acquisition conditions. Classification models for tumor detection are often assessed using several complementary key metrics. Accuracy measures the overall correctness of the classification (ie, distinguishing tumor vs no tumor cases); however, this measure alone can be misleading in highly imbalanced tasks, such as detecting low-incidence tumors. Sensitivity, or recall, evaluates the ability of a model to detect cases with tumors (true positive rate), while specificity quantifies its ability to correctly identify negative controls (true negative rate).
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - While any form of tumor detection can be valuable, models that perform well only on large, hard-to-miss tumors offer minimal benefit to the diagnostic workflow. Despite advancements in automated tumor diagnosis using DL models, detecting small pancreatic tumors in medical imaging remains one of the greatest challenges. Small lesions often exhibit only subtle visual cues that can be easily overlooked, even by experienced radiologists.Although the definition of a small tumor can be somewhat arbitrary, it is most commonly defined as one with the largest diameter of less than 2 centimeters (ie T1 stage). This threshold has demonstrated prognostic significance, with higher chances of tumor resectability and lower risk of metastasis at presentation.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Despite significant advancements in DL models for pancreatic tumor detection, several challenges remain before widespread clinical implementation can be achieved. A major challenge models face in achieving clinical applicability is generalizing performance beyond the training dataset. For instance, Liu et al. observed a significant performance decline in their CNN classification model, trained exclusively on an Asian population, when evaluated on an external cohort from the United States. They attributed the performance drop to differences in population characteristics and scanning parameters, highlighting the variability in model robustness across diverse patient groups.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - To address these challenges, standardization initiatives such as CLAIM (Checklist for Artificial Intelligence in Medical Imaging) aim to establish guidelines for model development, validation, and reporting, ensuring greater consistency across AI models.47,48By harmonizing evaluation protocols and promoting transparent reporting, these initiatives help bridge the gap between research and clinical implementation of DL-based tumor detection models. Other efforts, such as openly sharing code and final model versions alongside published results, further enhance model evaluation, reproducibility, and continuous model improvement.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - To address this challenge, ongoing efforts in Explainable AI focus on techniques such as saliency maps, feature attribution and attention mechanisms, which help visualize the most influential regions within an image that drive the prediction.50By enhancing transparency, these methods improve interpretability and foster clinician trust in DL-based diagnostic models.51Integrating explainability frameworks, improving cross-cohort model validation, and aligning AI models with standardized reporting guidelines will be essential to ensure reproducibility, robustness, and move toward the widespread clinical adoption of automated pancreatic tumor screening.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - One promising direction in multimodal detection is the analysis of longitudinal electronic health records (EHR). Emerging applications of LLMs show strong potential to extract information from unstructured clinical narratives and temporal trends, generating representations that can be interpreted by AI systems.These models can potentially identify subtle and evolving signals, such as symptom patterns, laboratory value changes, or medication adjustments, that may appear before anatomical abnormalities become detectable. For instance, in the setting of pancreatic cancer, the onset of new or worsening diabetes has been recognized as a potential early marker of disease.AI models may be able to detect concerning patterns in advance and integrate them with imaging and molecular data into multimodal prediction frameworks, enabling earlier risk stratification and more effective diagnostic and monitoring strategies.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Another promising area of research is the integration of CT detection models with liquid biopsy techniques. Liquid biopsy provides a minimally invasive approach to detect tumor-derived material in blood or other body fluids, including circulating tumor DNA (ctDNA), circulating tumor cells, and exosomal markers. Although more commonly used for disease monitoring, these biomarkers have also demonstrated potential for identifying early molecular changes related to pancreatic cancer, often before abnormalities can be seen on imaging studies.4Among these, ctDNA has garnered significant attention due to its ability to detect clonal somatic mutations, such as the KRAS mutations commonly found in pancreatic tumors.Studies have shown that combining ctDNA analysis with protein biomarkers (eg, CA19-9, CEA, HGF, and osteopontin) can significantly improve diagnostic sensitivity for pancreatic cancer.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - The future of early pancreatic cancer detection will depend on developing multimodal AI systems that are trained on rich and diverse sources of information, encompassing imaging, clinical records, molecular profiling, and systemic biomarkers. As these technologies evolve, rigorous validation, adherence to regulatory standards, and seamless integration into existing radiology workflows will be necessary to build trust among physicians and fully realize the potential of AI-driven pancreatic tumor detection in clinical practice.
Early detection of pancreatic cancer on computed tomography: advancements with deep learning.
Lopez-Ramirez F, Syailendra EA, Tixier F, Kawamoto S, Fishman EK, Chu LC.
Radiol Adv. 2025 Aug 19;2(5) - Question What are the associations of CT-evaluated body composition with early recurrence and overall survival in patients with pancreatic ductal adenocarcinoma after resection?
Findings High visceral-to-subcutaneous fat ratio is an independent predictor for early recurrence, whereas high skeletal muscle density and subcutaneous fat area independently predict better overall survival.
Prognostic value of body composition on early recurrence and long-term survival of resectable pancreatic ductal adenocarcinoma
Linxia Wu, Tong Nie, Xiaoling Zhi et al.
European Radiology 2025 (in press) https://doi.org/10.1007/s00330-025-12028-8 - Results High VSR was an independent predictor for ER (OR: 2.304, p=0.001) and worse OS (HR: 1.462, p=0.007), whereas high SMD (HR: 0.609, p=0.005) and high SFA (HR: 0.649, p=0.002) were independent predictors for better OS. Subgroup analyses revealed variations in the prognostic effect of VSR according to diabetes status and tumor size. A model combining body composition metrics and clinicopathological indicators (carbohydrate antigen 19-9, carbohydrate antigen 12-5, tumor-node-metastasis stage, lymphovascular invasion, and adjuvant therapy) demonstrated good predictive ability for ER, with AUCs of 0.80 in the training set and 0.82 in the validation set.
Conclusion High VSR was an independent predictor for ER and worse OS in PDAC. Moreover, combining bod composition metrics and clinicopathological indicators can improvetheprognosis prediction of patients with PDAC after surgery.
Prognostic value of body composition on early recurrence and long-term survival of resectable pancreatic ductal adenocarcinoma
Linxia Wu, Tong Nie, Xiaoling Zhi et al.
European Radiology 2025 (in press) https://doi.org/10.1007/s00330-025-12028-8 - This study highlighted the utility and prognostic value of preoperative CT-based body composition in predicting ER and overall survival in patients with PDAC. The integrated model combining body composition parameters with clinicopathological indicators, including CA 19-9, CA 12-5, TNM stage, lymphovascular invasion, and adjuvant therapy, demonstrated good predictive performance for patient prognosis. These findings could help identify patients at a high risk of ER and assist clinicians in delivering personalized management and trpeatment strategies.
Prognostic value of body composition on early recurrence and long-term survival of resectable pancreatic ductal adenocarcinoma
Linxia Wu, Tong Nie, Xiaoling Zhi et al.
European Radiology 2025 (in press) https://doi.org/10.1007/s00330-025-12028-8 - Our findings show that AI trained on large, diverse datasets can exceed average radiologist performance in PDAC detection on routine contrast-enhanced CT, providing a foundation for regulatory dialogue and prospective validation. We further show the generalisability of these findings by validating AI performance on two independent external cohorts from different countries that were not used for model training. Future validation trials are required to further evaluate AI performance across broader clinical settings and patient populations, as well as to assess workflow integration and assisted reading scenarios. Ongoing in-depth analyses of AI and reader performance stratified by reader experience, tumour size, and clinical stage will be the subject of a forthcoming study aimed at clarifying how AI can most effectively support clinicians in practice
- Background Pancreatic ductal adenocarcinoma (PDAC) has the worst prognosis among major cancer types, primarily due to late diagnosis on contrast-enhanced CT. Artificial intelligence (AI) can improve diagnostic performance, but robust benchmarks and reliable comparison to radiologists’ performance are scarce. We established an open-source benchmark with the aim of investigating AI systems for PDAC detection on CT and compared them to radiologists’ performance, at scale.
Interpretation AI demonstrated substantially improved PDAC detection on routine CT scans compared to radiologists on average, showing potential to detect cancer earlier and improve patient outcomes.
Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study
Natalia Alves et al.
Lancet Oncology 2025 (in press) - Purpose This study investigates the frequency, progression, and clinical implications of pancreatic findings on chest low-dose computed tomography (LDCT) scans performed for lung cancer screening.
Conclusion Incidental pancreatic findings were uncommon (0.9%) and included calcifications, atrophy/fatty infiltration, cysts, ductal dilatation, and masses. These findings do not by themselves indicate pancreatic cancer but warrant documentation and, when suspicious, dedicated pancreatic imaging. Radiologist scrutiny could improve detection accuracy, indicating the potential of a LDCT lung cancer screening program for detecting and monitoring pancreatic lesions.
Pancreatic findings in participants in a program of low-dose computed tomography screening for lung cancer
Gros, Louisa,b; Yip, Rowenaa; Zhu, Yeqinga; Li, Pengfeia; Paksashvili, Natelaa; Sun, Qia; Yankelevitz, David F.a; Henschke, Claudia I.a
European Journal of Cancer Prevention ():10.1097/CEJ.0000000000000997, November 18, 2025 - Out of 9467 participants, 90 (0.9%) had pancreatic findings, mostly male (54.4%), median age 64.7, with smoking (92.2%), alcohol use (41.1%), and diabetes (22%). Of these, 60 (66.7%) were detected on baseline LDCT, primarily as calcifications (73.3%), atrophy/fatty infiltration (18.3%), and duct dilatation (5%). Of the 90 participants, 27 underwent only baseline LDCT. Among the remaining 63, 33 had pancreatic findings on baseline scans, 27 of whom (81.8%) showed consistent findings on follow-up, and 30 developed pancreatic findings during surveillance. Rereview of the baseline scans showed that 68 participants (75.6%) had findings, including eight missed earlier. More cases of atrophy/fatty infiltration and other findings were detected compared to the original report, with calcifications remaining predominant (50 participants). Similar patterns were observed during the rereview of the latest LDCT scans. Two participants with detected lesions underwent biopsy, diagnosing a serous cystadenoma and pancreatic adenocarcinoma. The latter succumbed to pancreatic cancer.
Pancreatic findings in participants in a program of low-dose computed tomography screening for lung cancer
Gros, Louisa,b; Yip, Rowenaa; Zhu, Yeqinga; Li, Pengfeia; Paksashvili, Natelaa; Sun, Qia; Yankelevitz, David F.a; Henschke, Claudia I.a
European Journal of Cancer Prevention ():10.1097/CEJ.0000000000000997, November 18, 2025
- Introduction: Three-dimensional (3D) reconstruction transforms cross-sectional medical images into interactive anatomical models, interpretable on an LCD screen, in augmented reality or via 3D printing. Although certain benefits have been established in liver surgery, its use in pancreatic surgery remains limited. This update outlines the applications of 3D visualization in pancreatic surgery, ranging from surgical planning to teaching.
Results: The analysis of these studies suggests that 3D reconstruction, in comparison to cross-sectional imaging, could improve preoperative evaluation, by facilitating the detection of anatomical variations, the assessment of resection margins, and the prediction of morbidity and mortality according to tumor volume and residual pancreatic parenchyma. 3D imaging could also improve intraoperative safety, with some series reporting a 50% reduction of blood loss and a 25% reduction in operating time. 3D reconstruction is also a promising tool for teaching surgical anatomy, particularly through 3D printing.
Conclusion: 3D reconstruction could improve outcomes of pancreatic surgery but requires robust comparative studies before becoming a standard evidence-based practice.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - Conversely, the use of 3D reconstructions in pancreatic surgery is still limited, and the scientific literature on the subject remains scarce. However, pancreatic surgery is also notorious for its complexity: resectability is determined by anatomic vascular relationships, the risk of bleeding is high,postoperative complications are frequent, and the prognosis of pancreatic disease, particularly malignancies, remains generally bleak with a low 5-year survival rate. These characteristics make 3D reconstructions a particularly relevant field, both to improve surgical planning and anticipate technical difficulties as well as to refine resectability criteria.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al. Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - 3D reconstruction could offer a substantial advantage over 2D imaging in the preoperative evaluation of resection margins, especially that of arterial resection, which isa key element in the surgical strategy of pancreatic tumor surgery. In a retrospective study of 105 patients, Griser et al. showed that the diagnostic performance (area under the curve of the receiver-operator characteristic curve) forthe assessment of arterial invasion was statistically significantly better than that of 2D. These results were confirmed by Fang et al., who, thanks to an improvement in sensitivity, specificity, and positive and negative predictive values provided by 3D reconstruction, proposed a new classification of tumor resectability based on these three-dimensional models. One of the hypotheses put forward toexplain this superiority is that 2D imaging tends to under-estimate the true tumor volume and that there is acorrelation between tumor volume, TNM stage and the risk of recurrence.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011- 3D reconstruction of the pancreas is currently based on 2D images (CT or MRI), the quality of which depends on the thickness of the slices and the stability of the patient. Man-ual or semi-automated reconstructions are time-consuming and require expertise. Outsourcing processing by specialized private companies or the development of automated tools,particularly through deep learning, could improve the accuracy and reduce processing time. Medico-economic studies comparing the different solutions should be conducted to move from a frenzy for innovation to evidence-based practices.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011 - To date, the traditional open laparotomy approach remains preferable in pancreatic surgery in many settings, particularly in the case of a locally advanced tumor or a complex surgical procedure. Regarding PD, which isthe most common intervention, randomized trials comparing laparoscopic versus the open route have not formally demonstrated a clear benefit of the minimally invasive approach in terms of morbidity and mortality or oncologi-cal outcomes . Similarly, the robotic approach, although innovative, still needs to prove itself in high-level methodology studies before being generalized.In this context, AR could represent a real game-changer for minimally invasive approaches. Indeed, the possibility of projecting 3D reconstructions in real time directly on the screen during laparoscopic or robotic-assisted surgery could compensate for the absence of tactile and visual cues that are specific to open surgery. This technological integration, which is difficult to transpose to conventional surgery,would strengthen the safety and precision of gestures in minimally invasive procedures.
Value of 3D reconstructions in pancreatic surgery: Current status
E. Roussel, J. Pinson, L. Duhamel et al.
Journal of Visceral Surgery, https://doi.org/10.1016/j.jviscsurg.2025.09.011
- Pancreatic ductal adenocarcinomas (PDACs) are often detected at an unresectable stage, leading to high rates of morbidity and mortality.However, patients with small PDACs (≤2 cm) have demonstrated better overall prognosis. PDACs are primarily evaluated and staged using a three-phase intravenous (IV) contrast-enhanced pancreas protocol computed tomography (CT) scan, which is also the key modality for determining local resectability. The overall sensitivity of CT for PDAC detection is about 90 %, but it drops to 63–77 % for PDAC ≤2 cm in size. Studies suggest that up to 40 % of small PDACs may go undetected on CT scans, highlighting the considerable challenges associated with identifying these tumors using conventional imaging methods, and representing a significant missed opportunity for potential curative surgery.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - In the context of medical imaging, path tracing applied to volumetric data offers physically based rendering for the visualization of anatomical structures. Rather than relying on explicitly segmented surfaces, this technique reconstructs the appearance of tissue boundaries and internal features by interpreting volumetric scalar fields — e.g., Hounsfield units in CT scans — as continuous spatial distributions of optical properties. A central component of this approach is the transfer function, which maps scalar intensity values to material-specific properties such as color and opacity. These mappings allow different tissue types to be differentiated based on their density profiles and enable the selective enhancement or suppression of anatomical structures. The transfer function thereby implicitly defines "surfaces" within the volume as regions of high opacity gradients or significant radiometric contribution.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - Path tracing operates by casting rays into the volume and stochastically integrating the radiative transfer equation along each path. At discrete intervals along a ray’s trajectory, the local scalar field is sampled, and the corresponding optical properties are retrieved from the transfer function. Since medical volumes are stored on discrete grids, interpolation — typically trilinear — is employed to estimate scalar values between voxel centers. This interpolation smooths transitions but can also introduce intermediate values in regions with high gradients, such as tissue interfaces. As a result, the optical properties in these transition zones may be significantly altered, potentially softening the visual appearance of otherwise sharp boundaries or affecting the perceived thickness and translucency of structures.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - Cinematic rendering offers an opportunity to improve the detection of small PDACs through its enhanced, photorealistic visualization. This technique highlights subtle textural differences between pancreatic tumors and the surrounding normal parenchyma, making it easier to identify small lesions that might otherwise be missed. CR can enhance the understanding of the spatial relationship between a mass and surrounding structures, thereby increasing diagnostic confidence In our practice, we have developed a series of presets for different organs or clinical applications, including eight specific presets designed for the pancreas. However, these presets can be manipulated in real time to accentuate the optical surface properties for improved tumor detection for specific cases.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - As discussed above, the ability of CR to enhance surface differences and internal architecture offers the potential for improved diagnostic accuracy. Fig. 3 depicts a pancreatic head tumor that is visualized on both CT and cinematic rendering. The CR images accentuate the tumor’s surface texture, and contrast adjustments provide improved visualization of its internal architecture, as shown in images C and D of Fig. 3. Similarly, Fig. 4 demonstrates how CR enhances visualization of tumor density, texture, and vascular anatomy in an incidentally detected PDAC.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - As illustrated above, cinematic rendering has the capacity to enhance diagnostic accuracy in PDAC detection, particularly helpful for small lesions that may be iso-attenuating or too subtle to be identified by conventional imaging. Its interactive display settings can be tailored to the anatomy of interest, optimizing visualization of pathological processes. This can be utilized to differentiate PDAC from mimicking pancreatic lesions, including pancreatitis, mass forming pancreatitis, pancreatic neuroendocrine tumor (PNET), solid pseudopapillary neoplasm (SPN) or metastasis from a primary tumor. The capability to modify display settings can emphasize structures with higher Hounsfield units, such as PNETs. The improved depth perception enables sharper visualization of the internal architecture of cystic pancreatic neoplasms, including fine septations and mural nodularity. This improved detail may aid in distinguishing between several types of cystic neoplasms . CR has been shown valuable in visualizing the mixed cystic-solid components and vascular spatial relationships in cases of SPN, including those with rupture. Furthermore, it assists in differentiating vessel stretching caused by a large SPN, from true vessel involvement seen in PDAC, thereby enhancing diagnostic confidence.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - The most significant barrier is the complex algorithm utilized by CR that requires dedicated software along with additional computational power, substantial network bandwidth, and specialized training; all of which may not be routinely available in most clinical settings. The post-processing involved in interactive rendering to tailor display parameters for each clinical indication and pathology demands both high expertise and real-time calculations for each manipulation, further requiring longer processing times. Preset selection and parameter optimization are crucial for accurate representation of the anatomy and pathology of interest. This requires experience with the CR software to optimize parameters accurately for lesion identification. Incorrect adjustment of these settings can either hinder visualization of critical findings or lead to diagnostic errors by ‘creating’ new lesions that may not be truly present, particularly due to shadowing effects inherent to the lighting model. This may lead to overcall of clinically insignificant lesions leading to increased healthcare costs for the patients. Although CR significantly enhances photorealism, its clinical utility remains theoretical, and it is still unclear whether this improvement leads to greater diagnostic accuracy.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5. - It is anticipated that the parameter adjustments for a particular pathology can be standardized and potentially automated in the future.Currently, rendering each case with CR requires approximately 5–7 min when performed by an experienced radiologist, whereas less experienced radiologists may require additional time. However, the integration of artificial intelligence has the potential to streamline this process and significantly reduce rendering times.AI can potentially automate preset selection and optimization of parameters based on both patient related factors and the organ of interest to enhance lesion detection and avoid human biases. Furthermore, with the advancement of radiomics and machine learning, CR images could serve as a valuable resource for extracting prognostic information from high-order imaging features.Integration of CR technology with augmented reality devices like Microsoft’s HoloLens offers new opportunities for immersive surgical planning and better inter-provider communication, enabling real-time visualization and interaction with radiological images.
Detection of small pancreatic ductal adenocarcinoma: The potential of cinematic rendering in earlier lesion detection.
Arshad H, Krueger S, Chu LC, Fishman EK.
Curr Probl Diagn Radiol. 2025 Oct 25:S0363-0188(25)00197-5.
- Small pancreatic ductal adenocarcinomas (PDACs) generally have a more favorable prognosis; however, up to 40% may be missed on CT scan. Cinematic rendering (CR) is a novel three-dimensional post-processing technique that accentuates subtle textural differences and holds the potential to improve detection of small tumors. It utilizes a complex global lighting model generating photorealistic images with enhanced visualization of shadows and texture based on volumetric scalar fields (Hounsfield units). While CR has been studied for pre-surgical planning and vascular anatomy assessment, its role in detecting small PDACs remains underexplored. This pictorial essay highlights the potential of CR in detecting small PDACs and differentiating surface textures, while also addressing its limitations and future directions.
- Cinematic rendering (CR) is a newer 3D rendering technique, first described for PDAC assessment in 2018, that creates photorealistic images using a global lighting model. This approach improves anatomical visualization and depth perception. CR enhances the visibility of subtle textural differences, improving tumor conspicuity compared to conventional 2D reconstruction, 3D VR or MIP techniques. Unlike standard VR, which relies on a single light source, CR employs a more complex global lighting model incorporating multiple light sources. This approach enables more accurate representation of shadows, textures, and enhanced depth of field in the images.
- A central component of this approach is the transfer function, which maps scalar intensity values to material-specific properties such as color and opacity. These mappings allow different tissue types to be differentiated based on their density profiles and enable the selective enhancement or suppression of anatomical structures. The transfer function thereby implicitly defines "surfaces" within the volume as regions of high opacity gradients or significant radiometric contribution.
- Surface-like features emerge where the accumulated opacity reaches a perceptually significant threshold, often coinciding with sharp transitions in the scalar field. At these locations, a local gradient is used to approximate the surface normal, enabling shading via physically based lighting models. The decision to terminate, scatter, or continue a ray is governed by Monte Carlo sampling. Rays encountering highly opaque or emissive regions may be absorbed or contribute significantly to image synthesis through direct or indirect illumination. In contrast, rays traversing low-opacity regions are likely to proceed further into the volume, accumulating color and opacity until a termination criterion is met. Optimizing these factors may result in lesion detection which may be missed on routine CT post-processing.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942- In 2023, a deep learning algorithm was trained on over 3000 noncontrast CT scans to identify PCs and seven subtypes of non- PC lesions, with an AUC over 0.98 for lesion detection when applied to a multicenter validation dataset of over 6,000 patients. The algorithm significantly outperformed human readers, even when readers were provided contrast-enhanced CT scans. When the algorithm diagnosis probabilities were used as an aid to human readers, their performance identifying PCs was significantly improved. The authors additionally showed an ability to identify PCs on noncontrast chest CTs. With further real-world validation, such tools may be a way to utilize common imaging studies obtained for non-PC reasons to inform the selection of individuals for further assessment.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - Over the past decade, there has been substantial interest in using ML and AI to enhance the diagnostic utility of pancreatic imaging studies. A 2022 study evaluated four ML classifiers trained on a dataset of volumetric pancreas segmentations from patients with PC with prediagnostic contrast-enhanced CT scans along with matched controls. The best-performing of these models (support vector machine) correctly classified 95% of prediagnostic and 90% of control CT scans in the test set and showed a higher discrimination performance than did two radiologist readers.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - The DNA- and RNA-based PancreaSeq test, developed by then University of Pittsburgh (Pittsburgh, Pennsylvania, USA), initially evaluated 74 PC-related genetic alterations to identify high-risk cysts. Next, the test was narrowed to assess alterations in gene or mRNA expression of KRAS, GNAS, BRAF, TP53, PRKACA/B, ALK, NTRK1/3, RET, SMAD4, and CEACAM5 to determine whether a cyst was neoplastic. PancreaSeq showed 82% sensitivity and100% specificity in detecting advanced HGD and early cancer in pancreatic cysts. Of note, the test provides information on cyst type and the risk of progression to high-grade dysplasia or cancer and is currently implemented in select institutes.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - Furthermore, screening regimens should not be considered independent of cost, especially in settings limited by patient resources and/or availability of equipment and expertise, therefore, the development of cost-effective approaches is important. To advance the field of PC early detection, large international groups such as the CAPS research program and the PRECEDE Consortium will be necessary to enroll enough patients to adequately power PC early detection studies. The efforts of these groups, along with the many researchers working to better understand the biology of PC, may drive a shift toward improved outcomes for this challenging disease.
Challenges of early detection of pancreatic cancer
Michael J. Shen, Arsia Jamali, and Bryson W. Kato
Clin Invest. 2025;135(20):e191942 - In 2010, the National Comprehensive Cancer Network (NCCN) introduced guidelines based on multidetector CT findings to classify localized PDAC into resectable, borderline resectable, and unresectable categories. Resectable PDAC refers to tumors that are deemed suitable for surgical resection with clear margins. In this category, the tumor has not invaded major blood vessels or distant organs beyond what can be safely removed. Borderline resectable and initially unresectable PDAC indicates tumors that have some involvement or encasement of nearby blood vessels, such as the superior mesenteric artery or portal vein. These tumors require neoadjuvant therapy (chemotherapy with or without radiation therapy) to facilitate successful resection by downstaging the tumor. Unresectable PDAC refers to tumors that have extensive involvement of nearby blood vessels or distant metastases.
ACR Appropriateness Criteria ® Screening, Locoregional Assessment, and Surveillance of Pancreatic Ductal Adenocarcinoma: 2025 Update
Expert Panel on Gastrointestinal Imaging: Alice Fung, et al
J Am Coll Radiol 2025;22:S610-S624. 
- Multiple studies suggest that contrast-enhanced CT may be able to detect suspicious findings before the final PDAC diagnosis. One such study by Higashi et al reports that unsuspected pancreatic cancer was most commonly detected radiographically as a small solid lesion on contrast-enhanced CT. Another study suggests that within the 3 to 6 months prior to diagnosis, a contrast-enhanced CT may be 86% sensitive in the identification of findings suspicious for PDAC. In addition, Toshima et al report that focal suspicious pancreatic abnormalities may be detected at least 1 year prior to a diagnostic CT.
ACR Appropriateness Criteria ® Screening, Locoregional Assessment, and Surveillance of Pancreatic Ductal Adenocarcinoma: 2025 Update
Expert Panel on Gastrointestinal Imaging: Alice Fung, et al
J Am Coll Radiol 2025;22:S610-S624. - A study by Higashi et al reports that a small solid lesion on contrastenhanced prediagnostic CT was the most common radiologic feature suggestive of preclinical PDAC with a median size of 7.5 mm. Singh et al find that a standard CT may be 86% sensitive during the 3 to 6 months before the formal diagnosis of PDAC when evaluating for a mass lesion, main duct dilation or narrowing/cutoff, common bile duct cutoff, extrapancreatic soft tissue, and vascular involvement. A study by Toshima et al suggests that focal pancreatic abnormalities may be found at least 1 year prior to a diagnostic CT with the most common findings being focal parenchymal atrophy, focal faint parenchymal enhancement, and focal main duct changes.
ACR Appropriateness Criteria ® Screening, Locoregional Assessment, and Surveillance of Pancreatic Ductal Adenocarcinoma: 2025 Update
Expert Panel on Gastrointestinal Imaging: Alice Fung, et al
J Am Coll Radiol 2025;22:S610-S624. - CT has the advantage over other imaging modalities by way of its superior spatial resolution. Pancreatic protocol CT, consisting of pancreatic and portal venous phases, has been shown to be 90% sensitive and 99% specific for detecting solid pancreatic neoplasms. However, its sensitivity decreases to 77% for lesions <2 cm. Astudy comparing screening modalities shows that endoscopic ultrasound (EUS) detected pancreatic abnormalities in 42% of subjects, MRI in 35%, and CT in 11% where the mean detected lesion size was 0.55 cm. The Pancreatic Cancer Early Detection (PRECEDE) Consortium, an international multispecialty group of pancreatic specialists, suggests that pancreatic protocol CT may serve as an alternative to MRI/ MR cholangiopancreatography (MRCP) for screening highrisk patients. The addition of a delayed or equilibrium phase to the typical CT pancreatic and portal venous phases may improve sensitivity for small PDAC and detection of liver lesions and provide prognostic information.
ACR Appropriateness Criteria ® Screening, Locoregional Assessment, and Surveillance of Pancreatic Ductal Adenocarcinoma: 2025 Update
Expert Panel on Gastrointestinal Imaging: Alice Fung, et al
J Am Coll Radiol 2025;22:S610-S624. - FDG-PET’s low specificity for PDAC is likely due to the increased FDG avidity of inflammatory processes and other malignancies. This low specificity has been shown to improve with FDG-PET/ CT to help differentiate benign pancreatic pathology from malignancy with several studies suggesting similar or superior diagnostic accuracy for PDAC when compared to contrast-enhanced CT and MRI. The use of maximum standardized uptake value (SUV max ) has also been shown to be helpful in differentiating benign from malignant lesions and offer prognostic information.
ACR Appropriateness Criteria ® Screening, Locoregional Assessment, and Surveillance of Pancreatic Ductal Adenocarcinoma: 2025 Update
Expert Panel on Gastrointestinal Imaging: Alice Fung, et al
J Am Coll Radiol 2025;22:S610-S624. - Purpose: Serous cystadenomas (SCAs), solid pseudopapillary neoplasms (SPNs), neuroendocrine neoplasms (NENs), and mucinous cystic neoplasms (MCNs) are pancreatic tumors that frequently develop calcifications. Identifying the presence and pattern of calcifications on unenhanced CT scans can significantly aid radiologists in differential diagnosis.
Conclusions: Approximately 30% of pancreatic tumors exhibit calcifications. Punctate intratumoral calcifications are more indicative of NENs, whereas coarse calcifications strongly suggest SCAs, influencing the differential diagnosis.
Intratumoral calcifications in pancreatic neoplasms on unenhanced CT: frequency and diagnostic implications
Riccardo De Robertis · Maria Chiara Brunese· Nicolò Cardobi et al.
Radiol Med. 2025 Nov 3. doi: 10.1007/s11547-025-02142-4. Online ahead of print. - The first aim of this study was to evaluate the incidence of calcifications and the pattern of calcifications among pancreatic NENs, SCAs, MCNs, and SPNs at unenhanced CT. Calcifications were found in 27.7% of the lesions and were more frequent among SPNs (68.2% of cases), followed by NENs (33.5%), SCAs (15.9%), and MCNs (7%). Overall, tumors containing calcifications were larger than non-calcified ones (43.6 Å} 30.3 vs. 32.1 Å} 20.4, p < 0.001), even though this difference was significant only for NENs (p < 0.001).
Intratumoral calcifications in pancreatic neoplasms on unenhanced CT: frequency and diagnostic implications
Riccardo De Robertis · Maria Chiara Brunese· Nicolò Cardobi et al.
Radiol Med. 2025 Nov 3. doi: 10.1007/s11547-025-02142-4. Online ahead of print. - The incidental detection of a pancreatic lesion at CT examinations performed for unrelated reasons may pose a relevant diagnostic dilemma. While uncommonly found, with a reported incidence of around 4% pancreatic incidentalomas harbor a significantly higher malignancy rate than those detected in other organs. Proper and timely characterization is necessary through MRI integrated with CT features. While the presumptive diagnosis of noncalcified tumors on unenhanced CT mainly relies on nonspecific findings, such as tumor attenuation, margins, and presence of ductal dilatation, this study demonstrated that the pattern of calcifications, which are on the opposite easily detected at unenhanced CT, may be helpful to address the differential diagnosis, in particular, to distinguish between NENs, SCAs, and other tumor histotypes. Current evidence indicates that serous cystadenoma (SCA) is correctly diagnosed in only 33% of cases, underscoring that the diagnosis of SCA is still challenging.
Intratumoral calcifications in pancreatic neoplasms on unenhanced CT: frequency and diagnostic implications
Riccardo De Robertis · Maria Chiara Brunese· Nicolò Cardobi et al.
Radiol Med. 2025 Nov 3. doi: 10.1007/s11547-025-02142-4. Online ahead of print.
Spleen
- While splenic rupture or hemorrhage is most often associated with trauma, a variety of non-traumatic conditions can also cause life-threatening rupture or hemorrhage that require urgent evaluation and management, yet these may not always be considered high on the differential diagnosis in the absence of trauma. Other nomenclatures associated with non-traumatic splenic ruptures include ‘spontaneous, ‘idiopathic’, or even ‘occult’ rupture. In many cases, the spleen harbors an underlying disease process that predisposes it to rupture without direct trauma, hence why it is commonly referred to as ‘atraumatic’ or ‘non-traumatic’ rupture of the spleen. These are further subcategorized as pathologic and non-pathologic (idiopathic).
CT of spontaneous atraumatic splenic rupture: etiologies and imaging findings.
Yasrab M, Rahmatullah ZF, Chu LC, Kawamoto S, Fishman EK.
Emerg Radiol. 2025 Sep 30. doi: 10.1007/s10140-025-02383-w. Epub ahead of print. - Epidemiologically, only 7% of atraumatic splenic ruptures are idiopathic, while the rest are due to one or more underlying etiological factors that include neoplastic processes, inflammatory or autoimmune disorders, viral infections, and hematological conditions. Patients typically present with vague to sharp abdominal pain and tenderness, nausea and vomiting, referred left shoulder pain (Kehr’s sign) that is seen in up to half of all patients, drop in hemoglobin, and sudden hemodynamic instability and shock in cases of more severe bleeding. The presence of splenomegaly, which has been seen in up to 55% of patients with atraumatic splenic rupture, and age above 40 are significantly associated with a higher mortality rate. Management can involve interventional radiological procedures such as splenic artery embolization or surgical intervention via laparotomy and splenectomy, in addition to addressing the underlying etiology.
CT of spontaneous atraumatic splenic rupture: etiologies and imaging findings.
Yasrab M, Rahmatullah ZF, Chu LC, Kawamoto S, Fishman EK.
Emerg Radiol. 2025 Sep 30. doi: 10.1007/s10140-025-02383-w. Epub ahead of print. - Spontaneous Splenic Rupture
Infectious and inflammatory etiologies
Vasculopathy and thromboembolism
Benign masses and malignancies - Aneurysms and pseudoaneurysms of the splenic artery make up nearly 70% of all visceral aneurysms. True aneurysms are most commonly idiopathic, but associated causes include portal hypertension, chronic liver disease, atherosclerosis, and acute or chronic pancreatitis. Atypical intraparenchymal pseudoaneurysms, which tend to be more saccular in morphology, are less frequent and almost always secondary to an underlying cause, commonly pancreatitis, iatrogenic injury, or infection. They pose a particularly high risk of rupture (up to 37%) and are nearly always fatal when untreated. Certain connective tissue disorders such as Ehlers- Danlos syndrome and Marfan syndrome can cause noninflammatory vasculopathy and involve the splenic vasculature, increasing the risk of rupture and bleeding. Rare instances of splenic involvement due to fibromuscular dysplasia have also been reported.
CT of spontaneous atraumatic splenic rupture: etiologies and imaging findings.
Yasrab M, Rahmatullah ZF, Chu LC, Kawamoto S, Fishman EK.
Emerg Radiol. 2025 Sep 30. doi: 10.1007/s10140-025-02383-w. Epub ahead of print. - Malignant hematologic disorders are the most common cause of spontaneous splenic rupture, including acute myelogenous leukemia (AML), acute lymphoblastic leukemia (ALL), chronic myeloid leukemia (CML), lymphoma (discussed in the next section), myeloproliferative disorders such as polycythemia vera or myelofibrosis, and myelodysplastic syndromes. The presence of splenomegaly and age above 40 are significantly associated with increased mortality when ruptured. While the pathophysiology remains unclear, infiltrative processes and infarction are thought to be responsible for spontaneous splenic rupture. Apart from an enlarged spleen, sites of splenic infarcts and perisplenic fluid collections can be appreciated in cases of spontaneous splenic rupture.
CT of spontaneous atraumatic splenic rupture: etiologies and imaging findings.
Yasrab M, Rahmatullah ZF, Chu LC, Kawamoto S, Fishman EK.
Emerg Radiol. 2025 Sep 30. doi: 10.1007/s10140-025-02383-w. Epub ahead of print. - Malignant masses can be primary or metastatic lesions. Lymphoid neoplasms are the most common primary malignant splenic neoplasms, including Hodgkin and non-Hodgkin lymphoma subtypes (Fig. 9). They often present as part of systemic disease commonly with associated adenopathy, and rarely as primary site of disease (less than 2% of all lymphomas). CT findings range from homogenous splenomegaly to solitary or multiple nodules or masses. Splenomegaly is present in two-thirds of patients and the lesions are frequently hypoenhancing. Another major primary splenic malignancy is angiosarcoma. Primary angiosarcomas of the spleen are rare (less than 5% of all angiosarcomas) and present as single or multiple complex masses or nodules with irregular borders in the background of splenomegaly. They are heterogeneously enhancing due to areas of necrosis and hemorrhage, and distant metastases may already be present at diagnosis due to the aggressive nature of this disease.
CT of spontaneous atraumatic splenic rupture: etiologies and imaging findings.
Yasrab M, Rahmatullah ZF, Chu LC, Kawamoto S, Fishman EK.
Emerg Radiol. 2025 Sep 30. doi: 10.1007/s10140-025-02383-w. Epub ahead of print.