| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Computer Applications |
1 From the Department of Radiology, Rm S-056, 300 Pasteur Dr, Stanford University School of Medicine, Stanford, CA 94305. Received July 20, 1998; revision requested September 24; revision received November 18; accepted January 11, 1999. Supported in part by National Institutes of Health grants 1R01 CA72023, 1P41 RR09784-01, LM 07033; the Society for Computed Body Tomography and Magnetic Resonance; the Packard Foundation; the Lucas Foundation; and the Phil N. Allen Trust. C.F.B. is a 1997 RSNA Scholar. C.K. is a 1998 GENDEX/RSNA Medical Student/Scholar Assistant. Address reprint requests to C.F.B. (e-mail: cfb@s-word.stanford.edu).
 |
Abstract |
|---|
 TOP Abstract IntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
MATERIALS AND METHODS: CT colonographic data in a patient with a normal colon were used as base data for insertion of digitally synthesized polyps. Forty such polyps (3.5, 5, 7, and 10 mm in diameter) were randomly inserted in four copies of the base data. Axial CT studies, volume-rendered virtual endoscopic movies, and studies from a three-dimensional mode termed "panoramic endoscopy" were reviewed blindly and independently by two radiologists.
RESULTS: Detection improved with increasing polyp size. Trends in sensitivity were dependent on whether all inserted lesions or only visible lesions were considered, because modes differed in how completely the colonic surface was depicted. For both reviewers and all polyps 7 mm or larger, panoramic endoscopy resulted in significantly greater sensitivity (90%) than did virtual endoscopy (68%, P = .014). For visible lesions only, the sensitivities were 85%, 81%, and 60% for one reader and 65%, 62%, and 28% for the other for virtual endoscopy, panoramic endoscopy, and axial CT, respectively. Three-dimensional displays were more sensitive than two-dimensional displays (P < .05).
CONCLUSION: The sensitivity of panoramic endoscopy is higher than that of virtual endoscopy, because the former displays more of the colonic surface. Higher sensitivities for three-dimensional displays may justify the additional computation and review time.
Index terms: Computed tomography (CT), image display and recording, 75.12115, 75.12117 • Computed tomography (CT), three-dimensional, 75.12117 • Computers, simulation • Colon, CT, 75.12111, 75.12115, 75.12117 • Colon, neoplasms, 75.3119
|
Introduction |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
A major obstacle to the implementation of CT colonography in routine clinical practice, however, is the need to view an enormous amount of imaging data, which often includes 200–350 axial sections for a supine study and may be double that if imaging is also performed with the patient in the prone position. There is continuing debate (6–10) about whether review of two-dimensional (2D) axial CT sections alone is adequate or whether some form of three-dimensional (3D) display is necessary to increase polyp detection to the highest level. In early studies, differentiation of the relative contributions of 2D and 3D viewing strategies was difficult, because both strategies were most often used together (3). Complicating this issue is the fact that 2D and 3D viewing or display modes are being developed at an increasing rate. To determine the most efficient method of image review for CT colonography and to validate new methods, the efficacy of each display mode must be rigorously determined under controlled conditions.
To establish a "ground-truth" model for the testing of CT colonographic display modes, we developed methods for electronic insertion of polyps into CT colonographic data obtained in patients enrolled in a clinical trial (11). The realism of the synthetic polyps has been validated in a blinded trial in which radiologists experienced with CT colonography reviewed sets of axial CT sections and 3D endoscopic volume renderings of synthetic lesions and true lesions. The reviewers were unable to distinguish synthetic lesions from true lesions (11).
In this report, we describe our initial experience with the synthetic polyp model system to independently evaluate three display modes for CT colonography. Experimental data sets that contained sets of synthetic polyps were generated and reviewed by radiologists in axial CT mode, virtual endoscopic mode with forward- and reverse-viewing fly-through volume-rendered movies, and panoramic endoscopic mode, a recently developed 3D rendering method for depiction of the topography of the inner colonic surface as a flattened structure. Our purpose was to determine the sensitivity of radiologist observers for the detection of colonic polyps with each of these data review (display) modes.
|
MATERIALS AND METHODS |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Colonic cleansing was performed with a solution of polyethylene glycol and electrolytes (GoLYTELY; Braintree Laboratories, Braintree, Mass). After informed consent was obtained, air insufflation of the colon was performed, followed by CT with the patient in the supine position. A CT HiSpeed Advantage scanner (GE Medical Systems, Milwaukee, Wis) was used, with the following parameters: 120 kVp, 240 mA, 3-mm collimation, and 6 mm/sec table speed (pitch of 2.0). Scans were reconstructed at 1.0-mm intervals with a display field-of-view of 36 cm by using the manufacturer's standard reconstruction algorithm. There was essentially no retained water or fecal material present on the CT colonographic study. Reconstructed CT images of this patient's colon were used for each of the four test cases. In this way, identical "background" colonic morphology was provided for each case, and the only differences between cases were in the locations and sizes of inserted polyps.
Polyps with a simulated diameter of 3.5, 5, 7, or 10 mm were created as spherical objects by using a software CT simulator initially developed elsewhere and modified in our laboratory. The polyp diameters were chosen to represent approximate factor-of-two differences in cross-sectional area between the size categories: If the cross-sectional area is equal to 
r2, where r is the radius, and the diameter is 10 mm, then the cross-sectional area is 78.5 mm2; if the diameter is 7 mm, the area is 38.5 mm2; if the diameter is 5 mm, the area 19.6 mm2; and if the diameter is 3.5 mm, the area is 9.6 mm2.
The polyps were inserted along the inner aspect of the colonic wall according to methods described in detail elsewhere (11). Insertion locations were randomly selected in the subvolume of data that represented the colonic surface as follows: First, a random point along a path describing the centerline of the colon was chosen (12). Second, a random radial ray was cast from the path point to the nearby colonic surface, and the intersection of the ray with the wall represented the insertion point. When added to the base data, the initially spherical synthetic polyps became hemispheric, and approximately half their diameter protruded into the lumen. The portion of the polyp that overlapped the colonic wall itself or adjacent abdominal fat did not alter the attenuation or noise in the base CT data, owing to the use of nonlinear compositing methods. These methods also allowed the addition of polyps to thin haustral folds, with the resultant lesion protruding from only one side of the mucosal surface of the fold. Although we did not quantitatively assess the depth of insertion, we did inspect each polyp by using axial and focused virtual endoscopic views to verify that the lesions were indeed visible as protrusions from the colonic surface.
A total of 40 synthetic polyps were added to the four copies of the base data by randomly distributing lesions according to size category and location. Specifically, 10 lesions were added to each of two copies of the base data, nine were added to one copy, and 11 were added to the fourth copy. No changes in lesion location had to be made because of retained fluid in the base data, and lesions that were randomly assigned to be inserted on a haustral fold were not moved. In addition to these experimental data sets, we created "answer" data sets by adding 10-mm-diameter synthetic lesions with high attenuation (1,000 HU) to four additional copies of the base data. The answer lesions were inserted at exactly the same locations as the experimental synthetic lesions and were used by the study coordinator (C.K.) in the data analysis phase (described subsequently). These data were not available to the radiologist reviewers during the blinded portion of the study.
Three-dimensional Rendering
Rendering and image review were performed with Infinite Reality or O2 workstations (Silicon Graphics, Mountain View, Calif) equipped with 256–384 Mbyte of random access memory. Rendering engine software (VOXEL VIEW, version 2.5.1; Vital Images, Minneapolis, Minn) was used for 3D volume renderings and for review of axial CT sections in stack, or cine, mode. Central-path–planning software (12,13) that we developed was used for segmentation of the air-containing colon and generation of a central flight path that was used to create key-frame animations with the rendering engine, similar to the process described by Rubin et al (14).
Forward- and reverse-viewing volume-rendered endoscopic movies were created by using a 60° camera field of view. A diagram of the geometry involved is shown in Figure 1. Volume rendering parameters were similar to those being used by groups (14–17) using these techniques for virtual colonoscopy. These parameters include specification of opacity tables to render air transparent, application of a color table to render the colon in a pinkish hue, and application of a lighting model to simulate the reflectiveness of the inner colonic lining. Volume rendering parameters were kept fixed throughout the study. For each forward- or reverse-viewing movie, an image frame was created every 3 mm along the central path of the colon, which resulted in generation of 400 frames for each movie.
|
The third display mode we tested was a recently developed mode termed "panoramic endoscopy." The geometry for this mode is illustrated in Figure 2. The central path through the colon was used as the location of a 60°–field-of-view virtual camera, with the camera viewing direction perpendicular to the local direction of the path. The camera was rotated around the path in 60° increments, which generated six image panels. When these image panels were displayed side by side, they depicted a panoramic view of the colonic wall around the current location. Sequential panoramic frames were generated by moving the camera location along the path in 3-mm increments, with 360° (six by 60°) panoramic rendering at each point. The overall number of panoramic frames per case (400 frames) was designed to be similar to the number of images reviewed in an axial CT study and in the forward- and reverse-viewing endoscopic movies. Typically, preprocessing and rendering of the movies for panoramic endoscopy took 1
hours per case.
|
Each case was reviewed and scored for each of the three display modes: axial CT, virtual endoscopy, and panoramic endoscopy. Sets of axial images or animations for each case were randomized and presented to the reviewers on separate days. For example, during a given session, a reviewer might score axial CT sections from case 1, virtual endoscopic mode animations from case 2, and panoramic endoscopic mode animations from case 3. This mixing of modes and cases was performed to remove the ability to correlate findings from one display mode with those from another and to minimize case memory. Cases were reviewed during 1 month, with approximately 1 week between review sessions.
Axial CT images were viewed in a cine, or stack, mode on the workstation, with an image size of approximately 5 x 6 inches. All images were reviewed at window and level settings of 800 and -100 HU, respectively. Reviewers were allowed to adjust interactively the window and level settings, but they typically did not modify the settings. For each test data set, a total of 355 axial CT sections were reviewed.
Virtual endoscopic movies were reviewed by using MOVIEPLAYER software (Silicon Graphics) to interactively play animations to represent forward- and reverse-viewing movies. Reviewers could move forward or backward along the path of the colon but could not adjust rendering parameters or the viewing direction in these precomputed frames.
Panoramic endoscopic movies were reviewed similarly by using a customized MOVIEPLAYER interface. As with the virtual endoscopic mode, reviewers could move forward or backward along the path of the colon but could not otherwise adjust rendering parameters or viewing direction.
For each case and display mode, the reviewers recorded the spatial coordinates of each possible lesion and the review time. Estimates of lesion size were not solicited, because the study was designed to evaluate detection sensitivity and not necessarily lesion classification.
Analysis
The study coordinator was required to reconcile each reviewer's findings against truth on the bases of a thorough familiarity with the base data and knowledge of the exact spatial locations of inserted synthetic polyps. For axial CT study review, accurate identification of which synthetic lesions were detected and which were overlooked was relatively straightforward, because pixel coordinates for a reviewer's true-positive observations typically matched the insertion locations within a 2–3-pixel radius.
Reconciliation of the reviewers' results with the truth for the 3D modes was more challenging because the exact location and appearance of a synthetic lesion transformed from a 2D to a 3D representation was not necessarily intuitive. On a 3D image, one would expect that an inserted synthetic lesion should perturb the experimental image relative to a 3D image of the base data as generated from identical virtual-camera viewpoint and viewing directions.
For most of the 7- and 10-mm synthetic lesions, perturbations from matched 3D renderings of the base data were obvious. For some of the larger lesions and for many of the smaller lesions, however, precisely what had changed between the base and experimental 3D data sets was not immediately apparent, so further analysis was necessary. For this, we used the answer data sets, as described earlier, which contained 10-mm, high-attenuation synthetic lesions that were location-matched with the experimental polyps. By comparing 3D renderings of the answer data with identical renderings of the base data, we could determine precisely where a colonic surface perturbation would be expected on a matched rendering of the experimental data.
Once the exact 3D location was identified, a direct comparison of the experimental 3D image was made with the base 3D image to determine whether the experimental lesion was adequately represented to the reviewer in the precomputed animations. This process was very important for data analysis reasons, because for the 3D modes, not all of the inserted lesions were actually visible to the reviewers, owing to geometric or optical considerations involved in generation of the renderings. In the current implementation of the virtual endoscopic mode in particular, forward- and reverse-viewing movies created with a 60° camera field of view and the camera oriented toward the furthest visible point along the central path do not display the entirety of the colonic lining, as illustrated in Figure 3.
|
Statistical analysis involved pairwise comparisons of sensitivity levels between display modes or between reviewers. When comparing modes, the statistical power was, in some instances, increased by pooling observations across all polyp sizes or by pooling results obtained for the larger (7- and 10-mm) lesions. We have explicitly identified calculations based on such combinations. Because the sample sizes were relatively small, the most appropriate statistical model was one in which estimates of binomial proportions are used (18). Significant differences were obtained when the P value was less than or equal to .05.
Examples of Display Modes
Examples of the synthetic lesions used in the visualization trials are shown in Figures 4–6. Figure 4 shows axial CT scans of 3.5-mm and 10-mm synthetic polyps. In Figure 4a, a 3.5-mm lesion was randomly inserted on a haustral fold; in Figure 4b, a 10-mm lesion was inserted along the colonic wall. These magnified views illustrate the individual lesions but do not demonstrate that the reviewers' task was to identify such lesions within the context of the full field-of-view study of 355 CT sections. Figure 5 shows examples of virtual endoscopic images of 3.5-mm and 10-mm polyps. In this display mode, the reviewer's task was to identify polypoid lesions and to differentiate them from normal haustral folds. Figure 6 shows examples of 3.5- and 10-mm polyps on panoramic endoscopic images. As with the virtual endoscopic mode, the readers' task was to identify polypoid lesions and to differentiate them from haustral folds in the base data.
|
|
|
|
|
|
|
RESULTS |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
|
|
|
|
|
Figure 7a shows the detection sensitivity for reader 1 as a function of polyp size. A general trend of increasing detection sensitivity with increasing lesion size was present for each display mode. For the 3.5-, 5-, and 7-mm lesions, the following trend was observed: Sensitivity of panoramic endoscopy was greater than that of axial CT, which was greater than that of virtual endoscopy. Within each size category, differences between display modes were not statistically significant. For 10-mm lesions, the trend was as follows: Sensitivity of panoramic endoscopy was greater than that of virtual endoscopy, which was greater than that of axial CT. When we pooled results obtained with lesions of all sizes (Table 1) and tested for differences between the various pairs of display modes (eg, axial CT vs virtual endoscopy, virtual endoscopy vs panoramic endoscopy), there was no statistically significant difference in sensitivity between the three modes.
Analogous results for reader 2 are shown in Figure 7b. For detection sensitivity for reader 2, the trend was as follows: Sensitivity of panoramic endoscopy was greater than that of virtual endoscopy, which was greater than that of axial CT. When we pooled results across lesion sizes, the panoramic endoscopic and virtual endoscopic modes were significantly more sensitive than axial CT (P = .007 and .002, respectively), but the sensitivity of panoramic endoscopy was not significantly different from that of virtual endoscopy (P = .34).
Comparison of Figure 7a with Figure 7b and the pooled results in Table 1 shows that there were differences in sensitivity between the two reviewers despite identical image content and display methods for the test cases. When display modes were considered separately but results with polyps of all diameters were pooled (Table 1), the sensitivity of axial CT for reader 1 (60%) was significantly higher than that for reader 2 (28%; P = .001).
Interobserver differences were not significant when results obtained with the virtual endoscopic mode (55% vs 42% for reader 1 and reader 2, respectively; P = .26) or the panoramic endoscopic mode (75% vs 58%; P = .10) were compared. Overall, reader 1 detected 76 (63%) and reader 2 detected 51 (43%) of 120 lesions, for a significant difference owing to the discrepancy in sensitivity with the axial CT mode (P = .002).
Because of interobserver variability with axial CT scans, results for the two readers could not be combined for this mode. The two readers' results could, however, be combined for virtual endoscopy and panoramic endoscopy. This yielded a sensitivity of 66% (53 of 80 lesions) for panoramic endoscopy and 49% (39 of 80 lesions) for virtual endoscopy (Table 1). Although this result was statistically significant (P = .025), the overall sensitivity of 66% was not impressive. However, these results were obtained by pooling lesions of all sizes. By pooling results only from the lesions in the "clinically relevant" size categories of 7 and 10 mm (see Fig 7), the readers in combination detected 90% (36 of 40) of lesions at panoramic endoscopy and 68% (27 of 40) at virtual endoscopy, which was a statistically significant difference (P = .014).
Detection of Visible Lesions
As already described, the results shown in Figure 7 include influences on the readers due to different modes of data presentation and to "colonic surface coverage" with the 3D display modes. For each mode, the specific lesions that created a substantial perturbation in the base data were explicitly determined by means of careful comparison by the study coordinator of the experimental, base, and answer data sets. Table 2 shows the actual number of visible lesions for each display mode. Whereas all 40 inserted polyps were visible as distinct differences from the base data on axial CT studies, 26 inserted polyps were truly visible at virtual endoscopy, and 37 were visible at panoramic endoscopy.
|
Figure 8 illustrates the detection sensitivities of the different display modes for each reader when only the visible lesions were considered as potential observations. Table 3 summarizes the pooled sensitivity results for these visible lesions. By considering polyps of all sizes pooled together, the sensitivity of reader 1 (Fig 8a, Table 3) was essentially the same for virtual endoscopy (85%) and panoramic endoscopy (81%, P = .72). Both virtual endoscopy and panoramic endoscopy were significantly more sensitive than axial CT (60%) for the entire range of polyp sizes (P = .03 for both virtual endoscopy vs axial CT and panoramic endoscopy vs axial CT). Figure 8b shows analogous results for reader 2. As for reader 1, virtual endoscopy and panoramic endoscopy, with sensitivities of 65% and 62%, respectively, were significantly more sensitive than axial CT, which had a sensitivity of 28% (virtual endoscopy vs axial CT, P < .001; panoramic endoscopy vs axial CT, P = .001).
|
Review Time
Average case review times are given in Table 4. In general, the 3D studies took longer to review than did the axial CT studies. Mean review times were as follows: for axial CT studies, 8 minutes; for virtual endoscopic studies, 12.5 minutes; and for panoramic endoscopic studies, 13.2 minutes.
|
|
DISCUSSION |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
To date, there have not been rigorous trials in which different display modes of CT colonography were compared, to our knowledge. In most series (3,15,16), results from a combination of 2D and 3D display modes were compared with results from fiberoptic colonoscopy. When 2D and 3D modes are used together, it is difficult to establish independently the relative diagnostic contribution of each. Although results from initial investigations (7) suggested that 3D studies added incremental value over 2D studies, the authors of recent reports (6,8) have suggested that 2D studies alone may be sufficiently sensitive for initial data screening.
Whether a 2D, a 3D, or a hybrid approach will be the best initial screening method for CT colonographic data is important, because the effort involved in generating images for review is markedly different for 2D versus 3D modes. It ultimately will be crucial to determine if the extra effort required to produce 3D views is justified by the improved detection sensitivity.
A major confounding issue in clinical series is related to the reconciliation of CT colonographic results with those of the reference standard of fiberoptic colonoscopy. Operator dependence and inaccuracy in unique identification of neighboring lesions make reports of lesion-to-lesion correspondence between detection at CT colonography and detection at fiberoptic colonoscopy an educated guess, at best (21,22). Also limiting is the fact that relatively few clinically important lesions (
1 cm in diameter) are detected in screening patient populations. In one study (3), only 15 lesions in this size range were identified in 70 patients, and in another study (8), only six lesions with a diameter larger than 8 mm were found in 44 patients.
In this report, we have described systematic work designed to elucidate the most sensitive methods for initial review of CT colonographic data. By using a system of synthetic polyps inserted into CT data obtained in a patient, we have a ground-truth model with which we were able to exercise control over many potentially confounding experimental variables. Our motivation for developing such a model was similar to that of Seltzer and colleagues (23), who used synthetic pulmonary nodules to study display modes for chest CT.
In blinded trials, we assessed the sensitivity of two radiologists at review of studies obtained with axial CT, virtual endoscopy, and a recently developed display mode termed panoramic endoscopy. When all inserted synthetic lesions were considered, the results of reader 1 followed the trend of sensitivity of panoramic endoscopy greater than that of axial CT, which was greater than that of virtual endoscopy, without significant differences between the modes. For reader 2, the trend was of sensitivity of panoramic endoscopy greater than that of virtual endoscopy, which was greater than that of axial CT; for reader 2, the 3D modes resulted in significantly better sensitivity than did the 2D mode.
Differences in sensitivity trends for the two observers are largely explained in terms of different performances with axial CT studies: The sensitivity for reader 1 was significantly greater than that for reader 2 (Fig 7). We believe this finding highlights one of the major challenges with regard to interpretation of CT colonographic studies as a stack of axial sections as compared with interpretation of 3D studies, namely, that of directing the radiologist's attention to an area of interest. In an axial CT study, any given image may contain multiple colonic cross sections, so a systematic search of the colon is necessary to prevent overlooking lesions. Also, the relative size of polyps in a display can be very small as compared with the overall diameter of the abdomen (6). The relatively complex visual search required for axial CT scans is in contrast to that for 3D displays, in which the observer's visible scene is typically constrained to a specific segment of the colon. In the latter circumstance, even relatively subtle lesions may be detectable, as suggested by the trend toward improved detection of small (3.5- and 5-mm) lesions by reader 2 (Fig 8b).
When we asked how the review of axial sections was approached by each of our readers, reader 1 indicated that he traced the colon systematically, whereas reader 2 used a section-by-section approach, attempting to detect lesions on all sections of colon visible on a particular CT image. Given similar amounts of CT colonographic interpretation experience, these different search patterns for axial CT studies may account for the interobserver variability we encountered.
By analyzing the results as a function of all lesions inserted in the base data (Fig 7), we attempted to model a clinical CT colonographic study, in which polyps are distributed relatively uniformly along the colonic mucosa, without necessarily a tendency toward location either on or between haustral folds. (More general tendencies toward a larger number of distal vs proximal colonic lesions do occur but on a larger spatial scale [24].)
With 3D virtual endoscopy, it often is assumed that all of the colonic surface is presented for visualization and thus detection. In reality, polyps located around corners or behind haustral folds may not be within the visible portion of the colon, which makes detection of such polyps impossible (Fig 3). The combination of results from forward- and reverse-viewing directions helps minimize this problem but does not ensure 100% coverage (25). In the current study, the virtual endoscopic mode demonstrated 26 of the 40 (65%) inserted lesions, whereas 37 (92%) lesions were visible on panoramic endoscopic studies (Table 2). These findings help explain why detection sensitivity was significantly higher with panoramic endoscopy (66%) than with virtual endoscopy (49%) when results from the two readers were combined across all lesion sizes (Table 1). Moreover, the absolute sensitivity of panoramic endoscopy for lesions in the 7-and 10-mm categories (90%) was statistically higher than that of virtual endoscopy (68%).
On the basis of these results, we infer that virtual endoscopy as currently implemented has an upper limit in terms of sensitivity for polyp detection, because not all of the colonic surface is presented to the reader. As technology advances to the point that interactive or real-time volume rendering becomes possible, the sensitivity of virtual endoscopy could improve, because an observer would be able to move the virtual camera at will in the attempt to maximize surface display.
By analyzing the detection results only for the lesions that were potentially visible in the renderings shown to the reviewers, we compensated for variations in colonic surface visibility between display modes. We found that the sensitivity trends for the two readers were similar (Fig 8), with no significant differences between panoramic endoscopy and virtual endoscopy. This suggests that the 3D modes depicted individual lesions equally well, despite differences in image appearance due to orientation of the virtual camera. However, the sensitivities of the 3D modes significantly exceeded that of 2D axial CT for both reviewers (Table 3). This suggests that presentation of the imaging data in a 3D format provided important visual cues that made polyp detection easier than that with 2D data.
The concept that the 3D modes effectively constrain the reader's visual field and direct a start-to-finish search through the colon was alluded to earlier. Another contribution to the improved effectiveness of 3D display may be that haustral folds are depicted as they appear in their natural state, which allowed quick identification of the folds as normal structures. The apparent limitations of 2D review alone might be surmountable with training or improvements in image display that help guide the observer through the colon. In addition, augmentation of axial images with coronal, sagittal, or multiplanar reformations could help improve detection at the expense of some additional review time.
Our study has several limitations. A first limitation was our use of synthetic polyps rather than true lesions. This provided maximal control over the variables of lesion size, shape, and location (11). Although we believe that these simulated lesions accurately model the primary visual cues of native polyps, it is difficult to control for all facets of visual perception that could affect lesion detection.
A second limitation was the use of a single patient's colon for the base data. One could argue that expert reviewers seeing the same colon 12 times would recognize its features. However, because the appearance of the colon varied substantially between the three display modes we tested, the reviewers saw similar colons on only four occasions, so this source of case memory was less of an influence than might be expected. We recognize, however, that each patient has unique colonic curvatures and haustral fold morphology and variable amounts of retained fluid that will affect lesion detection. For this basic trial, which focused on relative performances with display modes, however, we believed that the benefits of a single well-cleansed colon outweighed the disadvantages. The primary advantage was that the background morphology was identical between experimental cases; thus, results from the four cases could be combined without concern about the introduction of a confounding variable.
A third limitation was that we inserted a relatively large number of lesions (nine to 11) in each experimental data set—more lesions than would be expected in a clinical case. Because this was a direct comparison between display modes, this issue probably is less important than it would be in clinical trials, where a "needle-in-a-haystack" search must be performed for what might be one clinically important polyp. In such a setting, reviewer fatigue likely will affect detection sensitivity.
Fourth, we used only spherical simulated lesions, which became hemispheric when composited with the base data. From a practical standpoint, insertion of spherical lesions is easier than insertion of asymmetric lesions, which might require a defined orientation with respect to the colonic wall. From a theoretic standpoint, we did not want to introduce shape features as an additional visual detection cue.
Finally, it is reasonable to question the value of studying each of the modes in a completely blinded fashion. To achieve rigorous, independent results, the rationale behind this design is obvious. In practice, however, a combination of 2D and 3D display modes will likely be needed to achieve not only the sensitivity but also the specificity necessary for clinical CT colonography (2,9,10,19). Lesion specificity, which is concerned with whether a detected polypoid structure represents a true polyp, a focus of retained stool, or a normal structure such as the ileocecal valve, is a critical issue for the overall success of CT colonography but was not the focus of the current study. Future investigators should also systematically address the effect of false-positive observations among different display modes, which could influence patient care with CT colonography.
Ultimately, what we expect is that an initial rapid screening of the imaging data in either a 3D or a 2D mode will be performed to identify areas of concern; then, supplementary images of the area will be created to aid in further analysis of the suggestive areas and decisions about whether the patient should be referred for fiberoptic colonoscopy. Although the current results are suggestive of an incremental value associated with 3D display for initial lesion detection, these results must be considered in the context of our specialized display methods and relatively limited sample sizes. We intend that this work will serve as a useful conceptual framework for future studies in which visualization issues will be systematically addressed.
In summary, we applied a ground-truth model system in initial evaluations of display modes for CT colonography. For 2D axial CT, detection sensitivity differed significantly between the two reviewers, a finding that may be related to different search patterns used during review. When we controlled for true lesion visibility, both reviewers achieved significantly higher detection sensitivity with the 3D modes of virtual endoscopy and panoramic endoscopy than with axial CT, which suggests that 3D images contained visual cues that made lesion detection easier, although review of 3D images was more time-consuming. On a lesion-by-lesion basis, reviewers performed similarly with both virtual endoscopic and panoramic endoscopic modes. However, for all inserted lesions, our panoramic rendering technique significantly increased the sensitivity for polyp detection above that of a display mode in which a conventional virtual camera is used, by increasing the percentage of displayed surface area. Thus, when comparing different display modes, one must recognize that detection sensitivity can be substantially limited because of geometric and optical constraints on the virtual camera, factors that, a priori, limit colonic surface depiction. Such factors could affect polyp detection at clinical CT colonography.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Abbreviations: 2D = two-dimensional 3D = three-dimensional
Author contributions: Guarantor of integrity of entire study, C.F.B.; study concepts, C.F.B., C.K., S.N.; study design, C.F.B., R.B.J., S.N.; definition of intellectual content, C.F.B., S.N.; literature research, C.F.B., C.K.; experimental studies, D.S.P., C.F.B., R.B.J., C.K.; data acquisition, C.F.B., R.B.J., C.K.; data analysis, C.F.B., C.K.; statistical analysis, C.F.B.; manuscript preparation, C.F.B.; manuscript editing and review, all authors.
|
References |
|---|
|
TOPAbstractIntroductionMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Related Article
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
