Chest: Technical Aspects of CT Angiography

Brian S. Kuszyk and Elliot K. Fishman Introduction | Data editing | 3D Rendering Techniques | Volume Rendering: Implementation | Volume Rendering:Parameters | Display | Conclusion | References

Introduction

A firm understanding of fundamental principles underlying CT angiography including spiral CT acquisition, image processing, and image display are required in order to get consistently excellent results over a wide range of clinical applications. One of the most compelling advantages of CT angiography is the ability to provide all of the information which previously required two or more radiological studies which may, in the case of conventional angiography, be much more expensive than CT. Image processing techniques such as volume rendering enhance this ability, allowing the radiologist and clinician to interactively explore different aspects of the dataset to address many specific questions which impact patient management.

In this review we describe the fundamental concepts underlying data acquisition, image processing, and image display. We focus on a practical approach to optimization in each of these key areas which we have found provides fast, reliable, and accurate results in a busy radiology practice.

Spiral CT Acquisition

A fundamental concept of 3-D imaging for any application is that the quality and accuracy of the resulting images is ultimately limited by the quality and resolution of the dataset. The rapid developments in spiral CT technology over the past 6 years have resulted in scanning capabilities for volume data acquisition that provide unparalleled opportunities for nothing short of reinventing CT applications and protocols. Today's spiral scanners have come a long way from the earliest prototypes which acquired initially a 12 sec and soon thereafter 24 sec study. While the latest "state of the art" capabilities are continually being upgraded, the typical top of the line spiral scanner can acquire data with subsecond (.7-.75 sec) gantry rotation, supports a spiral length of 40-60 sec, and can acquire a second back-to-back spiral with only a 5 sec interscan delay. Scan parameters such as kVp and mAs are identical to those used with state of the art non-spiral protocols, typically operating in the range of 300 mAs. Data reconstruction times vary from 1-5 sec per slice, and can be performed as a background function without interfering with the scanner's ability to scan other patients. The newest scanners acquire slices faster and allow longer total imaging times, thereby providing the large volumes of very high resolution data which are ideal for 3-D imaging.

A single breathhold for each phase of acquisition (i.e., arterial and venous) provides optimal results. With proper coaching, we have found that over 95% of patients can perform a 30-40 second breathhold. Shallow breathing works well in patients who are obviously not able to perform an extended breathhold.



Contrast Injection Techniques

Optimal CT angiographic images requires the rapid intravenous injection of contrast material via a power injector at 3-4 cc/sec. Higher rates up to 5-7 cc/sec have been used in the literature but are not generally necessary in our experience. Clave et al1 found in phantom studies that a luminal attenuation of 150 HU gives optimal results for measuring of carotid stenosis - this level of enhancement is easily achieved with an injection rate of 3-4 cc/sec. We use nonionic contrast injected via a 20 or 22 gauge angiocatheter in the antecubital vein.

Proper timing from the start of the injection to the start of scanning is essential to ensure imaging at the time of peak intravascular enhancement. Although automated techniques such as SmartPrep (GE Medical Systems) or C.A.R.E. Bolus (Siemens Medical Systems) are available to measure the contrast circulation time in individual patients, we have found that the use of a empiric delay of 25-30 seconds for arterial imaging in the chest and abdomen and a delay of 60-70 seconds for venous imaging is faster and yields excellent results in most patients. In older patients or patients with evidence of decreased cardiac function we typically increase the arterial and venous phase delays by approximately 10 seconds. The volume of contrast used should be adequate to maintain maximal vascular opacification throughout the spiral acquisition - 120-150 cc of contrast is adequate for most applications.

Because oral contrast may obscure intravascular contrast and necessitate more extensive editing of the dataset, we use water is used instead of positive constrast agents for abdominal applications of CT angiography. While this may at first seem to be a liability in terms of evaluating the bowel, we have found that in practice water is very effective oral contrast agent which may be better than positive oral contrast agents.

Vessel Orientation

The orientation of the vessel of interest has an important impact on the accuracy and appearance of CT angiography. Because spiral CT can currently only reconstruct axial images, resolution in the axial plane is always higher than in the z direction (perpendicular to the axial plane). Wise et al2 have shown in phantom studies that artifical lumen eccentricity can be a significant problem for vessels which are not oriented perpendicular to the axial plane. This phenomenon is due to the anisotropic nature of the spiral CT data. Likewise, we have shown in phantom studies3 that CT angiography with volume rendering is significantly more accurate in assessing percent stenosis for vessels oriented perpendicular to the axial plane than for vessels with more in-plane orientations. The carotid arteries are particularly well suited to accurate measurement of luminal diameter due to their orientation, whereas renal arteries may be more difficult to accurately assess due to their more in-plane orientation.

Collimation and Table Speed

As a general rule, it is usually preferable to increase the pitch up to 2 rather than the collimation in order to achieve adequate coverage of the volume of interest. Polacin et al4 have shown that a pitch of 2 increases the effective slice thickness minimally compared to a pitch of 1 when 180 degree linear interpolation is employed. However, a pitch of 2 doubles the distance covered in the z direction. This favorable tradeoff should be exploited in most CT angiography applications to allow use of the narrowest collimation possible.

Image Reconstruction Interval

One of the principal advantages of spiral CT over conventional CT is the ability to reconstruct images at any interval required. Overlapping reconstructions have been shown to improve the quality of 3-D images5,6, and are routinely used for CT angiographic applications. We have found that an overlap of 50% is adequate for most applications, although we routinely reconstruct spiral CT datasets at 1 mm intervals for highest resolution imaging of small vessels such as the renal arteries. Images can be reconstructed at any chosen interval from 1 to 10 mm, with the primary limiting factors being the time necessary to reconstruct the dataset (from less than 1 sec to 10 sec per reconstructed image) and the size of the resulting dataset, which may often be larger than 100 megabytes in size.

While reconstructing at 1 mm intervals does provide optimal accuracy for 3-D CT angiography, it also presents number of practical problems. Computer technology is rapidly making issues of reconstruction time and storage capacity less significant. The problem of how to meticulously review hundreds of axial images in a timely fashion is also significant. We typically film only every 3rd image, but review the entire dataset at the computer workstation. Cine mode, multiplanar reconstructions, and interactive real-time volume rendering can all be valuable for reviewing large datasets quickly and completely. Many current computer workstations require that the entire dataset be constructed at the same interval; this fact makes the use of narrow reconstruction intervals only through the area of interest impractical if 3-D reconstructions are to be performed. It is best to store the raw data until acceptable 3-D reconstructions are performed, thereby allowing later reconstructions at different intervals or using different reconstruction algorithms if necessary. Once the original data is deleted, retrospective changes in the reconstruction technique are not possible.

Subsecond Spiral Scanning

Subsecond spiral scanning times increase the distance which can be covered with a single spiral scan and further reduces motion artifacts without a significant reduction in image quality7,8, and therefore should be used for all CT angiography if available.

Data Editing

Editing of the spiral CT dataset is commonly used to remove other high attenuation structures such as bone which may interfere with visualization of the intravascular contrast. The need for data editing varies with the specific clinical application and the 3-D rendering technique used. While editing is absolutely essential with maximum intensity projection, other rendering techniques such as surface rendering and volume rendering may require little or no editing for effective visualization in many applications. Nevertheless, editing is useful and in some cases absolutely essential regardless of the rendering technique, and is an important and challenging area of ongoing research.

Manual Editing

Manual editing typically involves the user drawing a region of interest around the structures to be included in or excluded from the 3-D image. This is typically performed on the axial source images and may be accelerated by the use of slab techniques9 which allow contiguous slices to be grouped together and edited as a unit. Manual editing is time consuming (typically 30-60 minutes per case), but it is flexible and can be effectively applied to virtually all clinical applications. It is also widely available on virtually all commercially available 3-D software packages. Manual editing of the axial source images is not generally required when using real-time volume rendering with clip planes.

Automated Editing

Segmentation is the division of an image into multiple areas or objects (called primitives) with distinct features, such as individual organs or tumor. Humans perform this task using a complex analysis of size, shape, intensity, location, texture, and proximity to surrounding structures. Performing this task automatically using a computer has proven to be an extremely difficult problem which continues to be an important area of ongoing research. There is continues to be no general computer segmentation algorithm that can be applied to all medical images or all regions of the body. Consequently, despite significant advances in image processing techniques, time consuming manual editing of the CT dataset by an expert is still often required for optimal visualization.

Automated editing applications are available for specific domains such as the lung10, abdomen9, and liver11. These applications use a variety of computer techniques to identify wanted and unwanted structures. In general, all such applications will fail in some circumstances. Therefore, an expert user must monitor the automated editing and there must be a means for the user to correct errors. We have shown that an automated bone editing algorithm can significantly reduce editing time compared to manual editing for abdominal applications with limited input9. The interactivity provided by real time rendering greatly facilitates image editing by allowing very fast and effective manual editing with simple editing tools.



Cut-Planes

Real-time volume rendering with clip plane editing provides a flexible means of interactively editing the actual 3-D image12. In this technique, user prescribed clip-planes which can be positioned at any orientation or depth within the 3-D volume are used to remove unwanted data, enabling the user to better visualize structures within the volume which would otherwise be obscured by overlying tissues. Multiple cut planes can be used as needed to allow optimal visualization from multiple orientations. This simple technique is flexible and fast - a diagnostically useful image can be created in literally seconds. The clip planes can also be used to pan through the data, creating images which combine the features of 3-D images and multiplanar reconstructions.



3-D Rendering Techniques

The large volume of data generated by modern spiral scanners challenge traditional methods for viewing radiological studies. Where a conventional CT study might have provided 4 sheets of images (12 images per sheet) which could easily be reviewed by a radiologist sitting in front of a light box, today's spiral scanners can generate hundreds of images which require many sheets of film to display. This problem has fueled the development of computer graphics workstations which allow the radiologist and clinician to interactively explore spiral CT datasets using a variety of display formats including standard axial slices, reconstructed slices in any plane, or high quality 3-D images. Three-dimensional images integrate large volumes of data into a form which may be easier to interpret and similar to other familiar studies as catheter angiograms.

Most clinical studies of 3-D imaging to date have used surface rendering or maximum intensity projection (MIP) techniques for generating three-dimensional (3-D) images from CT datasets13-17. While some studies have shown that these 3-D techniques can be useful in clinical applications, several investigators have found standard axial and/or multiplanar images to be more accurate than the MIP or surfaced rendered 3-D images for a variety of CT angiography applications, including the carotid arteries15, the renal arteries16, and the aortoiliac system17. Such mixed results highlight the inherent limitations of these 3-D rendering techniques. In order to speed image processing, both surface rendering and MIP ignore most of the available CT data and use very simple schemes to distinguish vessels from other tissues18. These compromises limit accuracy and are therefore less attractive with each successive generation of computing power. Nevertheless, surface rendering and MIP are widely available techniques and are clinically useful. We will discuss these rendering techniques briefly before providing a more detailed discussion of volume rendering which provides a more flexible and accurate solution for 3-D visualization of CT angiographic data.

Surface Rendering

Surface Rendering was one of the earliest methods for 3-D display, and is available in most commercially available 3-D medical imaging packages. In this method, each voxel within the data set is determined to be a part of or not a part of the object of interest, usually by comparing the voxel intensity to some threshold value, thereby defining the "surface" of the object. With the surface determined, the remainder of the data is discarded. Surface contours are typically modeled as a collection of polygons and displayed with surface shading. The resulting image is a simplified representation of a structure which may be very inaccurate, particularly if the surface is difficult to determine precisely as is often the case in medical imaging. By converting the data from a volume to a surface, a large portion of the data available is forfeited in exchange for faster, easier computation. While this can be an advantage by allowing real-time rendering and thereby enhancing user interactivity, the usefulness of surface rendered medical images is generally limited by their artifacts and poor accuracy.



Maximum Intensity Projection

Like surface rendering, MIP is also commonly available in commercial 3-D software packages and so has been extensively clinically evaluated, particularly with respect to its usefulness in creating angiographic images from CT and MRI data. The MIP algorithm evaluates each voxel along a line from the viewer's eye through the image and selects the maximum voxel value as the value of the corresponding display pixel. The resulting images are typically not displayed with surface shading or other devices to help the user appreciated the "depth" of the rendering, making three-dimensional relationships difficult to assess. If there is another high intensity material along the ray through a vessel (such as calcification) the displayed pixel intensity will only represent the calcification and will contain no information from the intravascular contrast. Selection of the highest pixel value also increases the background mean of the image, particularly in enhancing structures such as the kidney and liver, thereby decreasing the visibility of vessels in these structures. Volume averaging coupled with the MIP algorithm commonly leads to MIP artifacts: a string of beads appearance in MIP images of normal vessels passing obliquely through a volume. While MIP has a number of important artifacts and shortcomings, it has been studied extensively and usually does provide superior accuracy to surface rendering for CT angiography14.

Volume Rendering

Volume rendering19-31 is a more advanced and computer intensive 3-D rendering algorithm that can incorporate all of the relevant data into the resulting 3-D image and overcomes many of the problems seen with surface rendering and MIP. Volume rendering is well suited to a wide range of medical and nonmedical visualization tasks due to its flexibility: data can be displayed with varying levels of opacity, surface shading, and perspective depending on the demands of each specific task. Continuing advances in computer power have transformed volume rendering from what was once a somewhat cumbersome technique requiring computer resources which were not widely available into a technique which can now be done at real-time frame rates (5-10 frames per second) using relatively inexpensive workstations.

Volume rendering was originally conceived at LucasFilms in San Rafael, California. The computer graphics group there was created by Ed Catamull, Ph.D., and Alvey Ray Smith, Ph.D., who were recruited by George Lucas to develop new computer graphics techniques to create more realistic images for the movies. Early examples of their work included the special effects in the "Star Wars" and "Star Trek" movies. They developed their own parallel processing computer, the Pixar image computer, and used its speed as the basis for further advances. Volume rendering was developed by three team members: Robert Drebin, Pat Hanrahan, and Loren Carpenter21. Volume rendering was unique in that it was applicable to a wide range applications including seismic data display, wind tunnel testing, and medical imaging.

As the name implies, volume rendering renders the entire volume of data rather than just surfaces or maximum intensity voxels, and so potentially conveys more information than a surface model. Volume rendering techniques sum the contributions of each voxel along a line from the viewer's eye through the data set. Because the information from the entire data set is incorporated into the resulting image, much more powerful computers are necessary to do volume rendering at a reasonable speed. We view volume techniques as the most advanced form of 3-D rendering currently available for creating accurate, clinically useful medical images. Volume rendering is just now being incorporated into commercially available software packages - with general availability and continued increases in computer power, it will likely become the most important rendering technique for 3-D medical imaging.

Volume Rendering: Implementations

Ray-Tracing

The original volume rendering algorithm described by Drebin, Carpenter, and Hanrahan used ray tracing to construct the 3-D image21. A number of specific implementations have been developed based on ray tracing. The ray-tracing approach sequentially computes the values for each displayed pixel in the 3-D image by calculating a weighted sum of all voxels encountered along a line (or ray) projected from the chosen viewing perspective through the data volume. This process is repeated by projecting a new parallel ray through the data for each displayed pixel. In order to create a 512 x 512 3-D image, this technique requires 262,144 sequential ray calculations! Additional calculations are required to incorporate surface shading into the image. While this approach can be slow, it can be implemented on very basic computer platforms, including personal computers without specialized graphics hardware.

Schreiner et al32 have shown that different variations of the MIP algorithm can result in very different images. Similarly, although specific implementations of the volume rendering algorithm by various manufacturers share important fundamental features, differences in the interpolation algorithms and other features may produce very different results both in terms of image appearance and accuracy.

Hardware Accelerated Techniques

Specialized computer graphics hardware is now commercially available which allows all of the pixel values in a 3-D image to be computed in parallel rather than the serial approach typically used by ray tracing programs. This approach affords dramatic improvements in rendering speed - volume rendered 3-D images can be rendered at real-time rates (5-20 frames/sec). Real-time rendering allows true user interactivity with the dataset, making possible such complex applications as simulations of minimally invasive and surgical procedures.

The real-time volume rendering application which was used to create the 3-D images in this paper is a modified version of the "Volren" real-time volume rendering program. Volren is a texture mapping based rendering package which was developed by a team led by Brian Cabral at Silicon Graphics27. The system uses trilinear interpolation hardware to extract a parallel set of oblique slices from a three-dimensional dataset and uses rasterization and compositing hardware to combine the slices in a way that models the passage of light through a partially transparent, emissive, and possibly reflective medium. This approach is much faster than traditional ray-tracing methods, but is currently more limited in terms of surface shading options.



Volume Rendering: Parameters

Window width and level

Volume rendering typically segments data based on voxel attenuation. We use window width and level controls which are similar to those used for conventional axial display of CT images. While the window can be adjusted to standard settings used to display soft tissue, liver, bone, or lung, the real-time rendering system permits the user to interactively alter the window setting and instantly see the changes reflected in the displayed 3-D image. This interactivity allows the user to rapidly customize the display to specific cases with varying levels of contrast enhancement and to rapidly explore a variety of attenuation ranges.

The transfer function used with volume rendering segments the data based on voxel attenuation but, unlike thresholding, it accurately models the physical reality that many voxels are only fractionally composed of intravascular contrast (or other materials). A standardized approach to selecting this transfer function is needed in order to ensure accurate, reproducible results for such applications as measuring vascular stenoses. Different rendering parameters can alter the apparent diameter of the normal vessel and the stenotic segment. In a recent phantom study of CT angiography with volume rendering3, we demonstrated the accuracy of the following approach for selecting the transfer functions: We assumed that voxels with an attenuation equal to or greater than the nominal attenuation of the intravascular contrast were composed of 100% contrast. Those with an attenuation less than or equal to the wall of the phantom vessel were considered to contain 0% intravascular contrast. Voxels with values between those of the wall and the intravascular contrast were considered to be only partially filled with contrast and assigned a percent intravascular contrast between 0% and 100%. The measurements of % stenosis had a mean error of 2% for vessels oriented perpendicular to the axial plane such as the carotids, suggesting that this is a valid approach for choosing the segmentation transfer function.

In clinical practice, a similar effect would be achieved by measuring the attenuation of the intravascular contrast and the adjacent soft tissue, which would then serve as the top and bottom points of the transfer function ramp respectively. The presence of mural calcification would require a modification of the transfer function used in this study, with a second ramp with downward slope at higher levels to separate intravascular contrast from calcium.

Opacity

Opacity refers to the degree which structures close to the user obscure structures which are further away. Opacity can be varied from 0% to 100%. Higher opacity values produce an appearance similar to surface rendering which helps to clearly display complex 3-D relationships. Lower opacity values allow the user to "see through" structures, and can be very useful for such applications as seeing a free-floating thrombus within the lumen of a vein or evaluating bony abnormalities such as tumors which are located below the cortical surface.

While these properties of opacity are intuitive, varying the opacity also has a second, less intuitive but very important effect on the image: it changes the apparent size of objects. Higher opacity values make objects appear larger, while lower opacity values make them appear smaller. This property has important implications applications which rely on measurements, including measuring % stenosis from CT angiography data. In our recent phantom study of CT angiography with volume rendering3, we found that an opacity of 50% gave the most accurate measurements of vessel diameter. Other opacity values may show specific features of the data to better advantage, but may also give inaccurate measurements. Further investigation is needed is this area to better characterize the interaction of opacity and other display parameters and their effect on the accuracy of the resulting image.

Brightness

The brightness can be varied from 0 to 100%. Brightness affects the appearance of the image, but does not affect accuracy - unlike opacity, it does not alter the apparent diameter of rendered structures. Brightness settings are largely subjective based on the preferences of the individual user. We have found that a setting of 100% works well for nearly all applications.



Accuracy

Improved accuracy is the primary reason for using volume rendering rather than a simpler technique such as surface rendering or MIP. Volume rendering has been shown to be superior to surface rendering for such musculoskeletal applications as the detection of fracture gaps20. Preliminary phantom studies in our laboratory have shown volume rendering to be very accurate for quantifying vascular stenoses3. Similarly, in a recent study of patients with suspected renal artery stenosis, Johnson et al30 reported that volume rendering was extremely accurate in identifying stenoses of 50% or greater. Our phantom studies have also shown the potential for significant interobserver variability with this technique31. This variability stems largely from the tremendous flexibility of the volume rendering technique. We are currently developing display strategies to ensure consistent results between readers.

Display

Conventional Computer Display

The simpler display techniques which can be used to convey a 3-D effect with conventional computer monitors and hard copies include depth shading, obscuration, and lighting33. Depth shading simply involves making more "distant" structures appear darker than those "closer" to the observer. While such an effect is easily achieved even with modest computer resources, the utility of this technique by itself is very limited. Obscuration is another relatively simple display technique whereby structures close to the observer obscure the view of more distant structures. This property is closely related to the opacity level chosen by the user. There is a tradeoff between the depth perception afforded by obscuration by a relatively opaque structure and the common medical necessity of seeing through transparent superficial structures to appreciate deeper ones. Lighting models vary widely from very simple ones based on the orientation of a surface relative to a single, fixed light source to much more complex models that account for multiple sources and the light reflected off of other structures within the rendered object. These complex models require much more computer power with often marginal improvements in viewer understanding compared to simple lighting models. Depth shading, obscuration, and some type of lighting are commonly used in combination in most 3-D rendering packages.

Kinetic depth effect refers to the depth cues than can be provided by rotating an object. This commonly used tool can be done with modest computer hardware by precalculating images from multiple angles around a single axis and displaying the images in rapid succession to provide a cine loop animation. Interactivity allows the users to control or alter the image to suite his or her needs. A simple example is user control of direction and speed of image rotation in a cine loop display. Real time volume rendering provides a higher level of interactivity which can significantly enhance the 3-D ques by allowing the user to alter the perspective and display parameters in real time. Real time rendering has the significant advantage allowing viewing form any angle, without the constraints imposed by precalculated video loop display.

Stereoscopic Viewing

Stereoscopic display techniques convey perspective and depth cues by providing slightly different images to the left and right eyes. This effect can be achieved by slightly altering the perspective of alternating images and using shutter devices incorporated into viewing eye wear which open and close to alternate frames between the left and right eyes. This technology is relatively inexpensive and can provide a dramatic 3-D effect which can be helpful in understanding complex anatomy. Head motion parallax allows the viewer to see an object from different angles as his or her head moves with respect to the display. When combined with stereoscopic viewing and real-time rendering speeds, this technique can provide a very realistic portrayal of 3-D relationships. While stereoscopic displays are not yet in routine clinical use, preliminary experiments in our laboratory show that both radiologists and nonradiologists prefer the stereoscopic display to conventional displays. We routinely view 3-D medical images using both conventional and stereoscopic displays.

Conclusion

Careful attention to technical details in all phases of CT angiography including data acquisition, image processing, and image display is essential in order to consistently produce optimal vascular studies. A basic understanding of each of these steps helps the radiologist to tailor the examination to specific clinical problems and avoid potential pitfalls. With optimal technique, CT angiography can provide very accurate images which obviate the need for conventional angiography in many circumstances. Continuing advances in scanner and image processing technology promise to further enhance both the accuracy and the practicality of CT angiography.



References

1. Claves JL, Wise SW, Hopper KD, et al: Evaluation of contrast densities in the diagnosis of carotid stenosis by CT angiography. AJR 169:569-573, 1997

2. Wise SW, Hopper KD, Ten Have T, Schwartz T: Measuring carotid artery stenosis using CT angiography: the dilemma of artificial lumen eccentricity. AJR 170:919-923, 1998

3. Kuszyk BS, Heath DG, Johnson PT, Fishman EK. CT angiography with volume rendering: in vitro optimization and evaluation of accuracy in quantifying stenoses. AJR 168(3):79, 1997

4. Polacin A, Kalender WA, Marchal G: Evaluation of section sensitivity profiles and image noise in spiral CT. Radiology 185:29-35, 1992

5. Ney DR, Fishman EK, Kawashima A, Robertson DD, Scott WW. Comparison of helical and serial CT with regard to three-dimensional imaging of musculoskeletal anatomy. Radiology 1992; 185:865-869.

6. Brink JA, Lim JT, Wang G, Heiken JP, Deyoe DA, Vannier MW: Technical optimization of spiral CT for depiction of renal artery stenosis: in vitro analysis. Radiology 194:157-163, 1995

7. Herts BR, Baker ME, Davros WJ, et al. Helical CT of the abdomen: comparison of image quality between scan times of .75 and 1 sec per revolution. AJR 167:58-60, 1996

8. Fishman EK. High-resolution three-dimensional imaging from subsecond helical CT datasets: applications in vascular imaging. AJR 169:441-443, 1997

9. Fishman EK, Liang CC, Kuszyk BS, Davi SE, Heath DG, Hentschel D, Duffy SV, Gupta A: Automated bone editing algorithm for CT angiography: preliminary results. AJR 166:669-672, 1996

10. Ney DR, Kuhlman JE, Hruban RH, Ren H, Hutchins GM, Fishman EK: Three-dimensional CT - volumetric reconstruction and display of the bronchial tree. Invest Radiol 25:736-742, 1990

11. Gao L, Heath DG, Kuszyk BS, Fishman EK: Automatic liver segmentation technique for three-dimensional visualization of CT data. Radiology 201:359-364, 1996

12. Johnson PT, Heath DG, Kuszyk BS, Fishman EK: CT angiography with volume rendering: advantages and applications in splanchnic vascular imaging. Radiology 200:564-568, 1996

13. Rubin GD, Dake MD, Napel SA, McDonnell CH, Jeffrey RB Jr: Three-dimensional spiral CT angiography of the abdomen: initial clinical experience. Radiology 186:147-152, 1993

14. Rubin GD, Dake MD, Napel S, et al: Spiral CT of renal artery stenosis: comparison of three-dimensional rendering techniques. Radiology 190:181-189, 1994

15. Cumming MJ, Morrow IM: Carotid artery stenosis: a prospective comparison of CT angiography and conventional angiography. AJR 163:517-523, 1994

16. Galanski M, Prokop M, Chavan A, Schaefer CM, Jandeleit K, Nischelsky JE: Renal artery stenoses: spiral CT angiography. Radiology 189:185-193, 1993

17. Rieker O, Düber C, Neufang A, Pitton M, Schwenden F, Thelen M. CT angiography versus intraarterial digital subtraction angiography for assessment of aortoiliac occlusive disease. AJR 1997; 169:1133-1138.

18. Heath DG, Soyer P, Kuszyk BS, et al: Three-dimensional spiral CT during arterial portography: comparison of three rendering techniques. RadioGraphics 15:1001-1011, 1995

19. Fishman EK, Drebin RA, Magid D, et al: Volumetric rendering techniques: applications for three-dimensional imaging of the hip. Radiology 161:56-61, 1987

20. Fishman EK, Drebin RA, Hruban RH, et al: Three-dimensional reconstruction of the human body. AJR 18:53-59, 1988

21. Drebin RA, Carpenter L, Hanrahan P. Volume rendering. Comput Graphics 1988; 22:65-74.

22. Drebin RA, Magid D, Robertson DD, Fishman EK: Fidelity of three-dimensional CT

imaging for detecting fracture gaps. J Comput Assist Tomogr 13:487-489, 1989;

23. Ney DR, Drebin RA, Fishman EK, Magid D: Volumetric rendering of computed tomographic data: principles and techniques. IEEE Comput Graphics Appl 10:24-32, 1990

24. Fishman EK, Magid D, Ney DR, et al: Three-dimensional imaging: state of the art. Radiology 181:321-337, 1991

25. Woodhouse CE, Ney DR, Sitzmann JV, Fishman EK: Spiral computed tomography arterial portography with three-dimensional volumetric rendering for oncologic surgery planning. Invest Radiol 29:1031-1037, 1994

26. Kuszyk BS, Heath DG, Ney DR, et al: CT angiography with volume rendering: imaging findings. AJR 165:445-448, 1995

27. Cabral B, Cam N, and Foran J: Accelerated Volume Rendering and Tomographic Reconstruction Using Texture Mapping Hardware. In: ACM/IEEE Symposium on Volume Visualization. 1994. Washington, DC.

28. Rubin GD, Beaulieu CF, Argiro V, et al. Perspective volume rendering of CT and MR images: applications for endoscopic imaging. Radiology 199:321-330, 1996

29. Kuszyk BS, Heath DG, Bliss DF, Fishman EK: Skeletal 3-D CT: advantages of volume rendering over surface rendering. Skeletal Radiol 25:207-214, 1996

30. Johnson PT, Halpern EJ, Kuszyk BS, et al: CT angiography of renal artery stenosis: comparison of real-time volume rendering algorithm with a maximum intensity projection algorithm. Radiology 205(P):295, 1997

31. Ebert DS, Heath DG, Kuszyk BS, et al. Evaluating the potential and problems of three-dimensional computed tomography measurements of arterial stenosis. J Digital Imag 11(3):1-8, 1998 (in press)

32. Schreiner S, Paschal CB, Galloway RL: Comparison of projection algorithms used for the construction of maximum intensity images. J Comput Assist Tomogr 20:56-67, 1996

33. Kuszyk BS, Ney DR, Fishman EK: The current state of the art in 3D oncologic imaging: an overview. Int J Radiation Oncology Biol Phys 33:1029-1039, 1995

34. Wise SW, Hopper KD, Schwartz TA, Ten Have TR: Technical factors of CT angiography studied with a carotid artery phantom. AJNR 18:401-408, 1997

35. Dix JE, Evans AJ, Kallmes DF, Sobel AH, Phillips CD: Accuracy and precision of CT angiography in a model of carotid artery bifurcation stenosis. AJNR 18:409-415, 1997