Elliot K. Fishman MD, FACR, Professor of Radiology and Oncology, Johns Hopkins University School of Medicine
Unlike many new technologies that spend time looking for the ideal application or a problem to solve, CT was quickly accepted into the medical community with its introduction in the mid to late 1970's. Even with those early 10-12 mm thick sections both researchers and clinicians recognized the limitations of an axial display especially when complex anatomy is to be assessed. The concept of 3 dimensional imaging quickly became of value to the referring physicians especially in such applications as craniofacial surgery (i.e. congenital anomalies) and orthopedic trauma (i.e. acetabular trauma) (1-7). In those early days the use and availability of CT for 3D imaging was limited by a number of factors ranging from the quality of scan volumes (i.e. thick sections, wide interscan spacing, long scan times), limiting computing power (i.e. early Digital Equipment VAX computers for 3D rendering), and limited 3D reconstruction algorithms (i.e. early versions of shaded surface display). Despite these limitations pioneers including Vannier and Hermann were able to publish early innovative papers that showed the potential value of 3D imaging (1-2).
With the continued parallel evolution of CT and computer technology, the clinical role of 3D imaging has continued to evolve and re-invent itself. In this chapter we will review the current state of the art of 3D CT imaging and reconstruction as well as trends and future directions.
The evolution of computer technology was one of the technical and social highlights of the 1980's and 1990's. Whether it be in designing faster computer chips (i.e. Intel Corporation), computer networks (i.e. Cisco Systems), post processing software algorithms (i.e. Pixar Inc.), data storage (i.e. EMC Corp) or increased network bandwidth (i.e. AT&T) others were developing technologies that could be put to good use in medicine especially in the evolving field of medical imaging. This dual use of technology is clearly defined by the development of volume rendering by LucasFilms (San Rafael, CA) and later Pixar Inc. (Richmond, CA). The technique as described in a patent filing by Drebin, Carpenter and Hanrahan was initially developed for improving animation and realism in the motion picture industry. (8) However, it quickly became clear that this technology had uses in any field where large volumes of data needed to be interpreted and represented in a format where the information could be mined for optimal image display and analysis. Fields that could take advantage of this technology included the oil industry for analysis of seismic data, the Defense Department for analysis of satellite data and medical imaging for the analysis of medical data like CT. The whole idea of reviewing data, as a volume became known as volume visualization, a term coined by Alvy Ray Smith in the mid-1980's.
Applications using volume rendering began moving into the research and clinical arena by 1986 although at that time hardware constraints were the single biggest limitation. We initially did our 3D volume rendering on a Pixar Image computer which with its interface (Sun 3/160 workstation, Sun Microsystems (Mountain View, CA) was able to process 40 million instructions per second using custom designed CHAP processors. This came at a cost of nearly 200-250,000 dollars. Time per clinical study, which consisted of video loops of 84 frames, was initially 24 hours but soon could be done in under an hour. Developments over the next 15 years led to significant decreases in hardware costs while hardware capabilities increased nearly as quickly. For example, a new processor board from ATI Inc. is 300 times more powerful and a thousand times cheaper than the Pixar Image Computer. (9) Hardware has moved from large and expensive computer workstations like those developed by companies like Silicon Graphics in the late 1980's and 1990's to more of a PC-based environment with dedicated processor boards in the late 1990's and early 2000's.
Changes in software also began to take advantage of more generic software as Windows NT from Microsoft Inc. (Redmond, WA). Initial systems like the Pixar Image Computer used system specific microcode while UNIX soon became the focus of development especially in the era when Silicon Graphics was the dominant hardware provider. More recently movement has been away from UNIX to the Microsoft NT world. Whether NT remains dominate is open to debate with the possibility of Linux a potential due to its lower costs, open code and advanced capabilities for select applications. Numerous vendors are looking at Linux as a next software solution.
Our current workstation is the Leonardo (Siemens Medical Solutions, Malvern, PA) running InSpace software and a syngo environment (Siemens Medical Solutions). This package runs on a standard PC with dual processor Intel Pentium IV Xeon processors, Nvidia Quadro Class Graphics Adapter and supplemented with special accelerator boards for faster processing of large volume datasets. The volume accelerator is the VolumePro 1000 from TeraRecon (San Mateo, CA), which allows us to do real time display of datasets of up to 1600 slices without any drop in performance. Datasets of over 2000 slices can also be done but with some slight decrease in performance. On this system we use a graphical interface that provides real time interactivity whether we are in axial mode, multiplanar reconstruction, volume rendering, MIP or MiniIP evaluation. High quality mapping of any tissue type is possible and interactivity moves the process from a secondary interpretation mode to a primary display and interpretation mode. There are of course a number of other workstations available whether they are from full product line companies like GE Medical Systems (Advantage workstation), or workstation only companies like TeraRecon and its Aquarius workstation. A comparison of the various workstations and hardware and software solutions is beyond the scope of this chapter. However it is assumed that in order to do many of the techniques discussed in this chapter that you have a current state of the art workstation and are familiar with its use. More will be commented on in this chapter regarding our expectations of the minimal necessary hardware and software needed as well as in our discussion on workflow in Chapter xx.
One of the topics that has been controversial in the past is the role of 3D imaging in clinical practice. Many radiologists have often felt that 3D images were of little value to them but did have some value to the referring physician. These radiologists often commented on the relative lack of quality of the 3D images when compared to the initial source CT data. We would agree that in many cases this was true but usually was due to either the poor quality of the initial datasets for 3D rendering or due to the poor quality of the 3D workstation, its algorithm or lack of user experience in using the workstation. This concept of limited value of 3D imaging was also illustrated by the workflow process in most institutions. That is, the 3D imaging was done nearly exclusively by technologists with little if any radiologist input. Although some centers have designed dedicated 3D imaging labs with dedicated technologists this has been an exception and in most cases the 3D images were generated in a less than ideal scenario. Although I do believe that dedicated technologists can do a good job, I believe that in order for 3D imaging to reach the mainstream that radiologists will have to take primary responsibility for the entire process. In addition, it is our experience that a paradigm shift to where 3D mapping is part of the primary study interpretation that the radiologist will be addressing at the time of initial inspection of the CT dataset. This change will not be a simple one since it does require changes in practice delivery and workflow. Yet, I would be willing to go out on a limb and boldly suggest that within 3-5 years this will become the technique of choice for delivery of our CT services.
To make this prediction come true there are several important barriers that must be overcome. The first is a logistic one, which finds the number and location of workstations to be limited in most institutions. Although some of the newer 16 slice MDCT scanners come with sophisticated 3D post processing software these are limited in number and location. PACs workstations are designed to read axial CT scans but were not designed to analyze volume datasets. These systems are classically designed as mouse driven soft copy systems, which were fine in an axial slice, based CT world, but are underpowered for todays 400-1000 slice CT datasets. It is both too expensive and too cumbersome to duplicate the functions needed with 3D CT workstations as well as some of the functions with classic PACs system workstations. The need to merge these two systems into a truly integrated single workstation will both lower costs as well as accelerate the move of volume visualization into the mainstream. Further discussion of this brave new world will be discussed in the chapter xx on CT workflow and e-training.
In one of the refresher courses we presented at RSNA on 3D imaging we were quoted as saying that the quality of a 3D image was very dependent on the quality of the initial dataset. The exact quote was "garbage in, garbage out" which does not sound very scientific but is a good description clearly stating that unless a quality dataset is obtained it is impossible to get a good 3D image regardless of the 3D technique used. The specific protocols for different applications are addressed in detail in the various chapters in this book, however several general themes cannot be over emphasized. There are several key parameters that need to be optimized in order to be able to obtain the best 3D images. These factors may be arbitrarily defined as scan parameters (i.e. slice thickness, interscan spacing), timing of contrast injection and data acquisition in contrast enhanced studies, and several other core study design factors.
The key to high quality 3D imaging is the use of thin sections reconstructed at close interscan spacing. While our pre-spiral CT protocol was 4 mm slice collimation reconstructed at 3 mm intervals our typical protocols for a 4-slice MDCT is the use of 1mm collimators, 1.25 mm slice thickness, and reconstruction at 1.00 mm intervals. With 16-slice MDCT we use .75 mm collimation, .75 mm slice thickness, and images reconstructed at .5 mm intervals. Specific scan protocols in regard to kVp and mAs will vary between different scanners but the selection of parameters must balance image quality with minimizing patient dose from the study. An up to date listing of protocols can be found on our website, www.ctisus.com. In cases where IV contrast is used (cases of CT angiography or other studies including virtual cystoscopy) the timing and delivery of contrast material relative to scan acquisition is critical. Optimal timing of arterial or venous phase imaging is dependent on the proper acquisition parameters. The use of preset defined timing delays (i.e. 25 seconds for arterial phase, 55 seconds for venous phase), timing based on test bolus injections or computer triggered imaging delays (i.e. preset value of 150 HU in aorta to trigger abdominal scan) have all been advocated by different authors. A good rule is to select the best technique for your institution knowing that different techniques work best for different applications. Finally, it is critical to understand the interplay of reconstruction algorithms and the ability to create quality 3D images. For example, when studying bone pathology images with a high spatial frequency reconstruction algorithm are ideal for bone definition when axial slices alone are considered. However, when 3D imaging is done especially with volume rendering the use of a high spatial reconstruction algorithm may result in images with too much noise. We have found that in select cases images may need to be reconstructed twice to provide the optimal datasets for the entire case. Once the dataset is acquired it will be necessary to analyze the data using 3D rendering algorithms.
Once a quality dataset has been acquired the rendering technique is the most important technical determinant of 3D image quality in most circumstances. (8,10-11) The rendering technique is the computer algorithm used to transform conventional serial transaxial CT imaging data into simulated 3D images. There exist a number of different methods for doing this, which can be divided into two classes: thresholding, or surface-based (binary), techniques and percentage, or semitransparent (continuum) volume-based, techniques. This initial selection of rendering technique has great impact on the quality of the final images in any given 3D application (12-17) .
Either technique consists of three steps: volume formation, classification, and image projection. Volume formation consists of the actual acquisition of the imaging data, the stacking of the resultant data to form a volume, and some preprocessing that varies according to the specific technique. Typical preprocessing includes resizing (by interpolation or re-sampling) of each volume element (voxel), image smoothing, and data editing (e.g., removing the CT table on which the patient lies). The classification step consists of determining the types of issue (or other classifying quality) that are present in each voxel and is either binary or continuous in nature. In CT, most voxels can be classified into four basic types: fat, soft tissue, or bone, and contrast enhanced tissue. Other imaging modalities may yield different categories of classification. The final step consists of projecting the classified volume data in such a manner that an image representing a view of the 3D volume from a chosen viewing orientation is displayed to the user.
Most early 3D imaging involved the use of thresholding-based imaging techniques, since thresholding can easily produce a model of surfaces of objects within the volume even with limited computer power. For thresholding classification, each type of tissue to be classified is assigned two numbers: the low and high thresholds. For a voxel to be considered as containing that tissue, its signal must lie within the range defined by the low and high thresholds. Bone is usually assigned a low threshold around 100 HU and a high threshold of more than 3,000 HU (essentially the top of the scale for most CT datasets). (18-19)
To classify the volume, the value or signal intensity at each voxel is analyzed and compared with the low and high thresholds for each tissue. If the signal intensity falls between the high and low thresholds defined for a tissue, the voxel is considered to contain that type of tissue. If the signal intensity lies outside the defined thresholds, it is considered to not contain that tissue type. The defined ranges of thresholds for various tissue types should not overlap. This classification is binary; that is, it defines each voxel as containing either 100% or 0% of a given tissue type, but nothing in between. Each tissue type is assigned a color (and possibly a level of transparency). Once the volume has been classified, most thresholding-based algorithms will extract surfaces from the classified data. A surface is defined as a boundary between voxels of one tissue type and voxels of another tissue type. An image can then be generated by defining a viewing orientation, calculating which surfaces would be visible from such an orientation, and projecting the information into a 2D viewing plane. The display may be reflective, with a simulated light source, or self-luminous, both of which provide perspective and depth cues.
The thresholding technique of classification has a number of limitations, the single biggest being that voxels that represent volume averaging (mixed tissue interfaces) cannot be correctly classified. Volume averaging is produced when two or more different types of tissue are present in one voxel. Thus, in CT, a voxel encompassing the boundary of muscle and bone will contain a volume average of attenuation values for bone and soft tissue. All imaging modalities will produce voxels with volume averaging because voxels have a finite size. With the thresholding classification, it is expected that each volume element contain one and only one type of tissue.* It is thus incompatible with volume averaging and incorrectly classifies voxels that contain volume averaging. The effects of volume averaging appear in the greatest number at tissue-surface interfaces. Of the voxels along the periosteal surface of a bone, for instance, many average both bone and apposed soft tissue. This geometric reality makes the accurate imaging of surfaces by means of thresholding classification difficult. Ubiquitous volume averaging makes it difficult to define a set of the thresholds that will represent a particular surface as it is modified by anatomic variation and pathologic conditions. That one must pick a fixed threshold severely constrains this technique. The threshold that would approximate bone in a healthy patient, for instance, exceeds the attenuation values for markedly osteopenic bone, creating artificial "holes" in the data and the final image. The thresholding technique is also susceptible to noise introduced in the scan. A small amount of noise can modify attenuation values, creating a soft tissue voxel out of one that is actually mostly bone.
All of these disadvantages add up to a number of deleterious effects on the final image: artificial holes in structures, artificial contours representing voxel boundaries rather than true tissue interfaces, artificial fragments of structures floating in space, and artificial absence or exaggeration of detail such as bone fractures. The main advantage of thresholding-based imaging is its speed, since a comparatively small amount of computational power is needed to generate images in a reasonable amount of time.
From a clinical perspective the limitations of thresholding techniques is underscored when one recognizes that less than 10% of the actual image data is represented in the final image. Although most clinical applications with thresholding based techniques had been with skeletal applications the technique was also used with variable success in CT angiography for display of the aorta and branch vessels. We do not use thresholding technique in our practice today and in many ways discussion of this technique may soon be only is of historical note.
Volume rendering is another technique for 3D display of medical data that came into use in the late 1980's. Volume rendering has the advantage that it can display data without classifying it into rigid all or nothing categories as thresholding does. Volume rendering is most often combined with a method of classification termed "Percentage Classification." The key difference between thresholding classification and percentage classification is that, in thresholding, it is assumed that each voxel contains either all or none of a particular tissue type, and no mixtures of tissues. In percentage classification, it is assumed that a voxel can contain one or more tissue types and the amount of each tissue is a continuum between zero and one hundred percent. This allows percentage classification to more closely approximate true voxel content in voxels containing tissue mixtures, or volume averaging. Percentage classification involves examination of each voxel to determine the amounts (percentages) of each tissue type present in the voxel. The resultant classified volume data consist of voxels still representing the percentage of each tissue type initially present.
The most common method used to determine the percentage contents is probabilistic classification involving a trapezoidal approximation. This method for determining tissue-type percentages works well for CT data. For trapezoidal classification, each tissue type is assigned a nominal value range that, in theory, represents that tissue type exactly. A voxel with a signal within that nominal value range is considered to contain 100% of that tissue. Around this ideal nominal value range, another range is defined by choosing a high and low point representing attenuation values at which a voxel would contain none of the designated tissue. Voxels with signal intensities that lie between the 0% point and the corresponding 100% points are assigned a corresponding percentage between 0% and 100%. Thus, a voxel with signal intensity precisely halfway between the 0% and 100% points would be assigned 50% of that issue. A voxel with signal intensity three-fourths of the way toward the 100% point would be assigned 75% of that tissue. All values between the 0% and 100% points represent voxels in which volume averaging is present (i.e., more than one tissue is present). This trapezoidal classification models closely the actual volume averaging in CT voxels.
Once the data have been assigned percentages, they must be further processed to form a final image. Each tissue is assigned a color and transparency. Each voxel is assigned a color and transparency by taking a weighted sum of the percentage of each tissue present in the voxel and the color and transparency assigned to those tissues. A final image is produced by casting simulated rays of light through the volume containing the classified and colored voxels. As the simulated rays pass through a voxel the color and transparency of the voxel modulates the color of the ray. The final result is an image that can be displayed on a computer screen or film. Volume rendering requires more computer power than surface based techniques because each voxel in the dataset must be projected into an image, whereas, with a surface based technique only the surfaces need to be processed. The final generated images using volume rendering do not have many of the significant computer-generated artifacts, found in surface based/thresholded images. Computer-generated artifacts at best, tend to engender distrust of 3D images and, at worst, could lead to profound diagnostic or therapeutic errors. We believe the greater fidelity of volume rendering with percentage classification to the patient data justifies the additional computer power required.
In terms of clinical applications one of the principle advantages of volume rendering is the ability to vary the opacity values, which allow selection of specific tissue types in a rendered image. Opacity refers to the degree with which structures that appear close to the user obscure structures that seem further away. Opacity can be varied between 0 to 100%. High opacity values can make images that accentuate the surface detail and look similar to surface rendered images. A low opacity value allows the user to see through structures and is especially useful in looking at bone and soft tissue and its relationship to vascular structures. One potential pitfall with varying opacity is that it may change object size, which may be important when grading stenosis. For example, higher opacity values make objects appear larger, whereas lower opacity values make objects appear smaller. Caution then is critical when using volume rendering for quantitative measurements.
Over the last few years, the medical imaging community has embraced volume rendering for a wide variety of 3D imaging applications including CT angiography, Oncologic imaging,Virtual Imaging and Orthopedics . The increasing power of computer hardware (and a reduction in its cost) makes volume rendering the technique of choice for 3D medical imaging.
Another technique we routinely use for CT angiography to supplement volume rendering is maximum intensity projection technique. (20-21) This technique is much simpler in principle than volume rendering as it looks at the entire dataset and projects the brightest objects (highest Hounsfield units) present. That is, pixels are displayed with gray scale relative to voxel attenuation. MIP provides no depth cues and because the brightest structures in the image seem closest to you the technique cannot be used to define 3D relationships. MIP technique cannot define in detail soft tissue and so has limitations when looking at details of organs such as the pancreas or liver. MIP however is ideal for looking at vessels and may be especially valuable in organs where vessel mapping is needed but organ enhancement may be significant. Two examples are defining the hepatic arterial or venous anatomy as well as defining the renal arteries as they travel into the renal cortex. MIP does have select limitations including a string of beads artifact in small vessels coursing obliquely through the dataset, and the potential overestimation of degree of vessel stenosis especially when calcium is present. Calcified plaque may obscure regions of stenosis and result in either an overcalling of the presence of stenosis or even suggesting vessel occlusion.
MIP typically requires editing of the dataset to remove bone or vascular mapping can be obscured (i.e. the aorta could not be seen on an AP projection as the spine would hide it). With newer workstations editing is fairly rapid so this is less of a technical issue than it was several years ago. Another modification of these techniques uses slaps of data rather than the whole volume to display MIP images. This often eliminates the need for any significant editing. In our experiences slabs of 20-50 mm usually work well in the chest or abdomen.
One potential practical advantage of MIP over VRT is that the implementation of volume rendering algorithms may differ significantly between different vendors to the point that one may have difficulty telling if an image is in fact volume rendered. Results in the published literature cannot be widely assumed to be true with every implementation of VRT in practice. With MIP the images are more likely to look similar regardless of the workstation or the end user. Recent changes in MIP algorithms have improved MIP quality and the latter statement may soon not be true either. Classically because of the impressive flexibility of VRT it is has traditionally been harder to train radiologists or radiologic technologists to become expert in this technique. The flexibility of VRT can also result in errors especially when VRT is used for quantification (i.e. measure per cent stenosis). Training therefore is critical before implementing VRT in your practice. However, on the newer workstation the learning curve is no longer a barrier to training and implementation.
In addition to the rendering technique important aspects of any 3D system are the system functionality in regard to image display and analysis. The most important one is the use of real time interactivity in viewing the datasets. Classic 3D imaging usually presented the radiologist and referring physician with a preset selection of views around one or more axis. Interactive real time rendering at a minimum of 8-10 frames per second but ideally at 20 frames per second or more eliminates the need for this and the user can choose from an infinite number of displays and projections, in real time. This helps the user select the single best view for any case or application. In addition display of data in stereo as well as capabilities for fly thru’s and fly around's can prove very useful. Stereo displays are especially valuable for defining vessel relationships including orientation, displacement and vessel encasement. (22-24). This is especially valuable with volume rendering. Stereo display conveys perspective and depth cues by presenting two separate renderings from slightly different points of view to the left and right eye. This results in an immediate perception of depth owing to the inherent integrating capability of the brain (stereopsis). Image separation on a single computer display is achieved with left and right shutter devices incorporated into eyewear that open and close to alternate frames. The process requires specialized hardware and software although the cost of these systems is going down.
Fly thru's are especially useful in such applications as virtual colonoscopy for polyp detecting and for virtual bronchoscopy for determining either extent of tumor, design or placement of a stent or even for planning surgical reconstruction or radiation therapy. (25-26) This technology is also being applied to endoluminal evaluation of the bladder. (27) On a practical note, fly thru's are becoming more popular with virtual colonoscopy as the user interface becomes both more intuitive as well as easier to use. The potential limitation may be in the low time it takes for the radiologist to review such a study (10-30 minutes). Further discussion of the principles and use fly thru's is provided in chapter xx on virtual colonoscopy.
The ability to use color in 3D imaging has been available for a number of years and in our experience has been used with mixed success. Color initially was used more as either a marketing tool, or as a way to make the cover of Diagnostic Imaging magazine. Colors were often chosen more for effect than for clinical utility. The coloring of organs as different colors (i.e. spleen in green, liver in red, kidney in blue, etc.) may seem interesting to a lay audience but usually was done as a way of trying to overcome the lack of fidelity in the 3D rendering.
Today color is being used with more impact and with important implications. Color for fly thru's with virtual colonoscopy enhance the realism of the dataset and are now becoming standard. Similarly, fly thru's of the airway or bladder also have increased value when appropriate color schemes are used. We have also found that color can be used with vascular imaging to enhance the 3D detail and spatial relationships especially when images are put on film or as slides (or in a textbook). The careful use of color can accentuate pathology and detail when used correctly. We have also found that color may enhance the 3D effect on a dataset by the use of shadowing and changes in the lighting model.
Another application of color is when imaging patients with orthopedic hardware. We have found that using the color blue to color the metal implant allows us to obtain 3D images with minimal artifact or noise. Applications range from post acetabular fracture repair to spinal screws to total hip replacement.
Finally, the use of varying the lighting model can enhance images when applied in select applications. We have found this to be especially true when imaging the skin in craniofacial imaging or when looking at colonic folds.
The advent of 4 and then 8 and finally 16 slice helical CT has provided the impetus for many changes in CT applications and implementation of study protocols. The impact has especially been felt with the ability to acquire true volume datasets which can be visualized as a true volume in a three dimensional world. In order to take advantage of this revolution in CT the radiologist must develop not only an understanding of the technical details of 3D imaging and the available rendering algorithms but also a hands-on knowledge of how to use it in clinical practice. Many of the chapters in this book will address these applications and will focus on the changes that this new technology is providing us today. Equally exciting, however, is that these changes will continue to evolve in the near future. Our recommendation is to embrace these changes and move forward in this brave new world of imaging.
1. Marsh JL, Vannier MW. Surface imaging from computerized tomographic scans. Surgery 1983;94:159-165.
2. Herman GT, Liu HK. Display of three-dimensional information in †omputed tomography. J Comput Assist Tomogr 1977; 1:155-160.
3 Totty WG, Vannier MV. Complex musculoskeletal anatomy: analysis using three dimensional surface reconstruction. Radiology 1984: 160:173-177.
4. Pate D, Resnick D, Andre M, et al. Perspective: Three dimensional imaging of the musculoskeletal system. AJR 1986; 147:545-551
5. Burk DL Jr, Mears DC, Kennedy WH, Cooperstein LA, Herbert DL
Three dimensional computed tomography of acetabular fractures. Radiology 1985; 155:183-186.
6. Fishman EK, Drebin RA, Magid D et al. Volumetric rendering techniques: Applications for three-dimensional imaging of the hip
Radiology 1987; 163:737-738.
7. Fishman EK, Magid D, Drebin RA, Ney DR, Kuhlman JE. Three dimensional imaging and display of musculoskeletal anatomy.
J Comput Assist Tomogr 1988; 12:465-467.
8. DrebinRA, Carpenter L, Hanrahan P. Volume Rendering. Comput Graph 1988; 22:65-74.
9. Robert A. Drebin, ATI (personal communication)
10 Fishman EK, Magid D, Ney DR et al. Three Dimensional Imaging.
Radiology 1991; 181:321-337.
11. Levoy M. Display of surfaces from volume data. IEEE Comput Graph Appl 1988; 8:29-37.
12. Scott WW Jr, Fishman EK, Magid D. Optimal imaging of acetabular fractures. Radiology 1987; 11:1017-1020.
13. Magid D, Fishman EK. Imaging of musculoskeletal trauma in three dimensions. Radiol Clin North Am 1989; 27:945-956.
14. Calhoun PS, Kuszyk BS, Heath DG, Carley JC, Fishman EK Three-dimensional Volume Rendering of Spiral CT Data: Theory and Method. RadioGraphics 1999; 19:745-764.
15. Johnson PT, Heath DG, Kuszyk BS, Fishman EK. CT angiography with volume rendering: advantages and applications in splanchnic vascular imaging. Radiology 1996:200:564-568.
16. Fishman EK. CT Angiography: Clinical Applications in the Abdomen
RadioGraphics 2001; 21:S3-S16.
17. Kuszyk BS, Heath DG, Ney DR et al. CT Angiography with volume †rendering: imaging findings. AJR 1995; 165:445-448.
18. Magnusson M, Lenz R, Danielsson PE Evaluation of methods for †shaded surface display of CT volumes. Comput Med Imaging Graph 1991;15:247-256.
19. Kuszyk BS, Heath DG, Bliss DF, Fishman EK Skeletal 3-D CT: advantages of volume rendering over surface rendering. Skeletal Radiol 1996; 25:207-214.
20. Napel S, Marks MP, Rubin GD et al. CT Angiography with spiral CT and maximum intensity projection. Radiology 1992; 185:607-610.
21. Johnson PT, Halpern EJ , Kuszyk B, et al. Renal Artery Stenosis: CT Angiography -- comparison of real time volume rendering and maximum intensity projection algorithms. Radiology 1999;211;337-343.
22. Johnson PT, Heath DG, Duckwall JR, Fishman EK Enhanced Display of Vascular Anatomy with Stereoscopic Viewing. J Diagnostic Radiography and Imaging 1999; 2(1): 25-28.
23. Smith PA, Marshall FF, Urban BA, Heath DG, Fishman EK. Three- dimensional CT stereoscopic visualization of renal masses: impact on diagnosis and patient management. AJR 1997; 169:1331-1334.
24. Smith PA, Heath DG, Fishman EK Virtual angioscopy using spiral CT and real-time interactive volume-rendering techniques. J Comput Assist Tomogr 1998; 22:212-214.
25. AP Royster, HM Fenlon, PD Clarke, DP Nunes, and JT Ferrucci. CT colonoscopy of colorectal neoplasms: two-dimensional and three- dimensional virtual-reality techniques with colonoscopic correlation
AJR 1997;169:1237-1242.
26. EG McFarland and JA Brink. Helical CT colonography (virtual colonoscopy): the challenge that †exists between advancing technology and generalizability. AJR 1999;173:549-559.
27. Jeong Kon Kim, Jae Hong Ahn, Taehan Park, Han Jong Ahn, Chung Soo Kim, and Kyoung-Sik Cho. Virtual Cystoscopy of the Contrast Material Filled Bladder in Patients with Gross Hematuria.
AJR 2002;179:763-768.