Gastrointestinal - Learning Modules - CTisus.com CT Scanning

Learning Modules ❯ Gastrointestinal

Effect of CT Reconstruction Algorithm on Diagnostic Performance of Radiomics Signature: Prediction of Pathologic Tumor Grades in Pancreatic Neuroendocrine Tumor

Elliot K. Fishman M.D.
Johns Hopkins Hospital

INTRODUCTION

Pancreatic neuroendocrine tumor (PNET) is a rare pancreatic neoplasm.
The incidence of PNET has increased significantly, in part related to the widespread use of cross-sectional imaging, currently accounting for 3-5% of all pancreatic neoplasms.
Very high heterogeneity of PNETs makes standardization of therapeutic strategies and optimal management difficult.
Radical oncological surgery is not mandatory for all PNETS anymore, and may be avoided for indolent, small PNETs.
WHO grading system of PNET is based on assessment of proliferative activity in the tumor, and is critical for determining treatment and prognosis.

INTRODUCTION (2)

Recent research has shown that CT radiomics analysis and CT features can be used to predict WHO grade of PNETs, and assist the clinical decision-making for patients with PNET.
However, many factors may affect the results of radiomics analysis, such as reconstruction algorithms, kernels, tube currents, and slice thickness.
Iterative reconstruction (IR) allows to substantially reduce image noise compared to conventional filtered back projection (FBP) and is now widely used in CT image reconstruction to improve image quality or to reduce patient radiation dose, though their effect on diagnostic performance of radiomics signature is not known.

PURPOSES

To evaluate the effect of CT reconstruction algorithms (FBP vs. IR) on diagnostic performance of radiomics signature to distinguish grade 1 PNETs from grade 2 and 3 PNETs.
To determine whether additional conventional CT image features could help predict pathologic tumor grades of PNETs.

Material and Methods: Patients

This study was performed following study approval from our IRB. Informed consent was not obtained.
144 patients (74 male, 70 female, average age 59.1 ± 12.7 years) with dedicated pancreas protocol CTs from our pathology and radiology database were retrospectively evaluated.
All patients had surgical resection of the tumor.
94 patients with pathological diagnosis of WHO 2010 grade 1 PNET (47 male, 47 female, average age: 58.1 ± 12.7 years)
50 patients with pathological diagnosis of WHO 2010 Grade 2 or 3 PENT (27 male, 23 female, average age: 60.4 ± 13.3 years)

CT TECHNIQUE

The patients were scanned on a dual-source multidetector CT scanner (Siemens Somatom Definition FLASH, Siemens Healthineers).
100-120 mL of iohexol (Omnipaque, GE Healthcare) at an injection rate of 4-5 mL/sec.
Scan protocols were customized for each patient to minimize dose on the order of 120 kVp, effective mAs of 290, pitch of 0.6, collimation of 128 x 0.6 mm
Arterial phase timed by bolus tracking usually 25-30 seconds post injection, followed by 30 second delay for venous phase.
Arterial phase CT data were reconstructed in both FBP and IR with kernel of B20f and I26f, respectively. Venous phase CT data were reconstructed in IR only with kernel of I26f, therefore, not used for radiomics analysis.
Reconstruction thickness/increment were 0.75 mm/0.5 mm.

Image segmentation

Both arterial and venous phase images were used for segmentation.
The entire three-dimensional (3D) volume of the pancreas as well as PNET were manually segmented using commercially available software, VelocityAI (Varian Medical Systems, USA)
Image segmentation was performed by one of four trained researchers and verified by one of 3 board-certified abdominal radiologists.

Image Analysis (Radiomics)

Arterial phase CT data with two sets of reconstructions (conventional FBP and IR) were analyzed.
A total of 478 radiomics features from the segmented volume using venous phase images were extracted to express pancreas phenotype.
A: 14 first-order statistics of the volumetric CT intensities
B: 8 shape features of the target structure
C: 33 texture features from a gray-level co-occurrence matrix and a gray- level run-length matrix
D: 368 texture features from the 8 filtered volumes by wavelets (8)
Random forest was used for the classification of PNET grade 1 and 2/3
Each decision node was divided until a unique case (n=1) remained.
Ten-fold cross-validation was performed in this study with 10,000 trees for each cross-validation.
The prediction of the class of test dataset was determined by majority voting from the trained trees.

Sample Selection Process for Radiomics analysis

Radiomics Feature Extraction Processes

478 radiomics features used in this study

A total of 478 features were computed from four categories.

Radiomics and PNETS

CT imaging Features

Following conventional CT imaging features were also analyzed.

Tumor size and location (Head/uncinate, body, tail)
CT density of the tumor on arterial and venous phase
CT density of the normal pancreas, abdominal aorta, and portal vein (venous phase)
If the tumor is hypodense than normal pancreas, it was recorded.
Presence of calcification within the tumor
Main pancreatic duct (MPD) dilatation (> 3mm) upstream to the tumor
Atrophy of the pancreas upstream to the tumor
Exophytic if >50% of tumor volume protruding from pancreatic contour
Liver metastasis
Suspicious lymph nodes (enlarged, abnormal morphology or enhancement)
Vascular involvement (compression/narrowing, encasement, tumor thrombus)
Cystic tumor (well-defined area of fluid attenuation < 30 HU)
Necrotic tumor (ill-defined area(s) of unehancement, attenuation < 40 HU)

Results: Demographic of PNET cases

Results: Radiomics prediction of PNET grade using entire pancreas and tumor

When the entire pancreas and tumor were used for the classification of the grades
[Iterative reconstruction] Overall accuracy: 69.4%, Grade 1: 82.8%, Grade 2 or 3: 50.0%
[Filtered back-projection] Overall accuracy: 75.5%, Grade 1: 82.8%, Grade 2 or 3: 65.0%

Results: Radiomics prediction of PNET grade using tumor only

When only tumor were used for the classification of the grades
[Iterative reconstruction] Overall accuracy: 81.6%, Grade 1: 86.2%, Grade 2 or 3: 75.0%
[Filtered back-projection] Overall accuracy: 85.7%, Grade 1: 89.7%, Grade 2 or 3: 80.0%

Results: Radiomics analysis

When the entire pancreas and tumor were used for classification, accuracy of prediction of grade 1 PNETs are similar between IR and FBP, with accuracy of 82.8% for both algorithms.
However, prediction of grade 2 or 3 PNETs was lower with IR than FBP, with accuracy of 50.0% and 65.0%, respectively.
Overall accuracy using the entire pancreas and tumor was 75.5% for FBP, and 69.4% for IR.
When only the tumor was used for classification, accuracy of predicting Grade 1 PNETs was higher with FBP (89.7%) than IR (86.2%).
Prediction of grade 2 or 3 PNETs was also higher with FBP (80.0%) than IR (75.0%).
Overall accuracy using the tumor only was 85.7% for FBP and 81.6% for IR.

Example case: 57-year-old man with 2.1 cm PNET (WHO grade 3) in the body of the pancreas seen as a hypervascular mass in arterial phase contrast-enhanced CT. This tumor was misclassified as grade 1 (by both of iterative reconstruction & filtered back projection).

Radiomics and PNETS

Results: CT image characteristics of PNET (1)

Results: CT image characteristics of PNET (2)

Results: CT image characteristics of PNET (3)

Results: CT imaging characteristics

Some CT imaging features were useful for differentiating grade 1 and grade 2 or 3 PNETs.
Grade 2 or 3 PNTEs tended to be larger compared to Grade 1 PNETs.
Grade 2 or 3 PNETs tended to be hypoenhancing relative to normal pancreas, and ratio of tumor density to the normal pancreatic density (in both arterial and venous phases), ratio of tumor density to the aortic density (in venous phase), and ratio of tumor density to the portal venous density (in venous phase) were significantly lower in Grade 2 or 3 PNETs compared to Grade 1 PNETs.
Presence of calcifications, upstream MPD dilatation, liver and lymph node metastasis, and vascular involvement were more commonly observed in grade 2 or 3 PNETs than grade 1 PNETs.

Discussion

By combining all 478 features, CT radiomics features predicted Grade 1 PNETs and Grade 2 or 3 PNETs with accuracy of 85.7% with FBP and 81.6% with IR.
Four cases of grade 1 PNETs were falsely classified as grade 2 or 3 with IR, and 3 cases of grade 1 PNETs were falsely classified as grade 2 or 3 with FBP.
Five cases of grade 2 or 3 PNETs were falsely classified as grade 1 with IR, and 4 cases of grade 2 or 3 PNETs were falsely classified as grade 1.
All other 42 our of 49 test cases were correctly classified by radiomics with both IR and FBP when only tumor was used for classification.

Limitations

Retrospective study, and radiomics signature was only evaluated with arterial phase CT because two sets of different reconstruction algorithms were not available in venous phase in our data.
Difference of tumor size between grade 1 PNETs and grade 2 or 3 PNETs may have influenced the prediction of grading of PNETs, although the tumor size is same between the two reconstruction sets.
Unable to assess radiomics signature or conventional CT features to differentiate grade 2 and grade 3 PNETs because of limited number of grade 3 cases were available.

Conclusion

Radiomics signature better predicted grade 1 PNETs and grade 2 or 3 PNETs using CT images reconstructed in FBP than those reconstructed in IR.
Diagnostic performance of radiomics signature is affected by CT reconstruction algorithm. Radiomics analysis is more accurate using CT images reconstructed in FBP than those reconstructed by IR to classify tumor grade of PNETs.
Some conventional CT imaging features were useful for differentiating grade 1 PNETs and grade 2 or 3 PNETs. Combining radiomics signature and conventional imaging features may improve prediction of tumor grade in PNETs.