| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Computer Applications |
1 From the Diagnostic Radiology Department, Warren Grant Magnuson Clinical Center, National Institutes of Health, Bldg 10, Rm 1C660, 10 Center Dr, MSC 1182, Bethesda, MD 20892-1182 (R.M.S., L.M.P., J.D.M., A.M.Y.); and the Department of Radiology, Mayo Clinic, Rochester, Minn (C.D.J., J.E.R.). Received May 5, 2000; revision requested June 13; revision received August 1; accepted September 12. Supported by the intramural research programs of the Diagnostic Radiology Department, Magnuson Clinical Center. Supported in part by NIH grant RO1CA75333. Address correspondence to R.M.S. (e-mail: rms@nih.gov).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
MATERIALS AND METHODS: Twenty patients with known polyps underwent CT colonography in the supine position. CT colonographic data were processed by using a shape-based algorithm that depicts masses that protrude into the lumen. We studied nine shape criteria and three isosurface threshold settings. Results were compared with those of conventional colonoscopy performed the same day.
RESULTS: There were 50 polyps (28 were 
10 mm in size; 12, 5–9 mm; 10, <5 mm). The sensitivity with optimal settings for detecting polyps 10 mm or greater was 64% (18 of 28). Sensitivity improved to 71% (10 of 14) for polyps 10 mm or greater in well-distended colonic segments. Performance decreased for polyps less than 10 mm, poorly distended colonic segments, and other shape algorithms. There was a mean of six false-positive lesion sites per colon. These sites were reduced 39% to 3.5 per colon by sampling CT attenuation at the lesion site and discarding sites having attenuation less than a threshold.
CONCLUSION: Automated detection of colonic polyps, especially clinically important large polyps, is feasible. Colonic distention is an important determinant of sensitivity. Further increases in sensitivity may be achieved by adding prone CT colonography.
Index terms: Colon, CT, 75.12115, 75.12119, 75.1282 • Colon neoplasms, diagnosis, 75.12115, 75.12119, 75.1282, 75.311 • Computed tomography (CT), computer programs • Computed tomography (CT), image processing • Computed tomography (CT), three-dimensional, 75.12117
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Computed tomographic (CT) colonography is a newer, relatively noninvasive colonic evaluation that may have higher patient acceptance (2). CT colonography is performed by first cleansing and insufflating the large bowel with room air or CO2 and then performing helical CT of the abdomen and pelvis. The examination can be performed in a few minutes without sedation or instrumentation, aside from placement of a rectal tube. CT colonography is presently under investigation at a number of research institutions and has been used to detect colonic polyps and cancers (3–8).
Although technical errors are being overcome, perceptual errors during the interpretation of CT colonographic findings will remain a problem and an important cause of false-negative examination results. Furthermore, CT colonographic interpretation is time-consuming; interpretation times of 15 minutes to 1 hour per patient have been reported (4,9). Time-consuming interpretation will add to the cost of the examination (10). Methods that reduce cost and improve diagnostic performance would be important improvements in this exciting newer technique.
Computer-assisted diagnosis has great potential to improve the efficiency and accuracy of interpretation of CT colonographic results. A promising algorithm for computer-assisted polyp detection has previously been shown to be successful in a CT colonographic colon phantom (11). The purpose of the current study was to improve the algorithm and test its feasibility in a selected patient population.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
CT Scanning
Patients underwent standard colonoscopic preparation with 1 gallon (3,785 mL) of an orally administered polyethylene glycol electrolyte solution (Colyte; Reed and Carnrick, Jersey City, NJ), two orally administered 5-mg bisacodyl tablets (Dulcolax; CIBA-Consumer, Edison, NJ), and CO2 insufflation of the colon to patient tolerance. Glucagon (1 mg) was intravenously administered to one patient to improve comfort; the remaining 19 patients did not receive glucagon. CT was performed with a helical scanner (HiSpeed Advantage; GE Medical Systems, Milwaukee, Wis). CT scanning parameters were 120 kVp, 70 mAs, a field of view to fit (38–46 cm), 5-mm collimation, a 1.3 helical pitch, and a 3-mm reconstruction interval (12). The protocol required three or four 20-second breath holds. A 3-cm overlap between breath holds was used so that no anatomy was missed because of a variation in breath-holding effort. Multiple breath holds and overlapping sections were needed because of a scanner limitation on the extent of anatomic coverage given the scanning technique used.
Patients underwent scanning in both the prone and supine positions. Supine scanning was performed first and was followed by prone scanning. Only the supine study was analyzed with the computer algorithm to simplify data analysis. The size of the typical CT data set was approximately 80 MB, or 160 images.
Conventional Colonoscopy
Conventional colonoscopy was performed after CT colonography on the same day. Colonoscopic examinations were performed by experienced colonoscopists with more than 3 years training. The size and location of any polyps were identified in the colonoscopy report. Polyp sizes were determined by the colonoscopist at the time of the examination by using a probe or forceps for reference. The colonoscopic pullback was videotaped for later review. The sizes of 11 polyps were given qualitatively (eg, as "diminutive" or "large") in the colonoscopy report. These sizes were determined on the basis of our best estimate from a review of the pathology report or the videotape.
Human Observers
The supine CT colonoscopic studies were interpreted at random assignment on the day of the study by one of three radiologists (including C.D.J.) who were blinded to the results of conventional colonoscopy. The radiologists had a strong suspicion that the patients had polyps somewhere in the colon because they knew that the case mix consisted of 160 positive and 20 negative cases. The observers recorded the presence and location of suspected polyps. All three radiologists were board certified and had at least 10 years of experience. They had each previously interpreted a minimum of 50 CT colonoscopic examinations. The radiologists used research software developed at the Mayo Clinic (by J.E.R.) for image display and interpretation.
Generation of Colonic Surface
We transferred the supine CT images to an Indigo2 workstation with Maximum Impact graphics (Silicon Graphics, Mountain View, Calif) and produced a three-dimensional surface rendering of the colon by using our endoscopic research software package (11,13–15). In brief, voxels within the colonic lumen are identified by using "region growing." In region growing, a single voxel within the colonic lumen (the "seed") is manually selected, and then the computer identifies all voxels connected to the seed that have intensities less than a specified threshold. Surface renderings of the colon and polyps were smoothed for purposes of presentation (16). No attempt was made to correct for the 3 cm of overlap, because polyps could be missed if overlap was removed and because colon distention was inconsistent in the area of overlap, which was not uncommon.
For each CT colonoscopic data set, we produced three surface renderings by using different parameter settings. The parameters adjusted were (a) the window width and level for converting the data from 12 to 8 bits, (b) the region-growing threshold for seeding the colonic lumen, and (c) the isosurface threshold for generating the isosurface (Table 1). The purpose of these experiments was to evaluate the importance of determining the precise location of the colonic wall for computer-assisted diagnosis of polyps and was motivated by results from our prior phantom study (11).
|
Development of Detection Criteria
Polyps were detected by using prototypic automated polyp detector software (11). This polyp-detector software is used to identify regions of the colon wall that have an abnormal shape by measuring local variations in curvature. The curvature is computed by calculating partial derivatives of the CT data by using a convolution kernel, which also smooths the images. The kernel size is 9 x 9 x 3 voxels (approximately 7 x 7 x 9 mm), which is slightly larger than that used in reference 11 because of a difference in section interval (3 mm in the current study, as compared with 1 mm in the phantom study).
The polyp-detection algorithm can be thought of as operating in two passes. In the first pass, areas of the colonic wall that protrude inward and are rounded on all sides are considered "candidate" polyps. We call this pass the "primary criterion," and the technical description of this type of shape is known as "elliptic curvature of the peak subtype." Although the primary shape criterion was used to detect most polyps, in our experience it produced too many false-positive findings (11).
Therefore, in a second pass, additional shape criteria were added to the algorithm to make it more selective; that is, to shorten the list of potential polyps. These additional shape criteria, called "filters," are summarized in Table 2. The filters consisted of various combinations of three parameters that describe shape: (a) mean curvature (H, the mean on a per-vertex basis of the principal curvatures Kmax and Kmin, where Kmax is maximum principal curvature and Kmin is minimum principal curvature, expressed in per-centimeter units); (b) the dimensionless ratio sphericity (Kmin - Kmax)/H, where the bars over each variable indicate means over all vertices that compose the polyp; and (c) minimum polyp size, expressed in centimeters or the number of vertices. A greater mean curvature indicates a more sharply curved potential polyp. Sphericity indicates how round the polyp appears and ranges from 0 to 2. A perfect sphere has a sphericity of 0. At the opposite extreme, a portion of the surface that is curved in one direction and almost flat in the perpendicular direction (approaching so-called cylindric curvature) has a sphericity of 2 because in that situation, Kmax 
0.
|
We tested filters 1–5 in five (25%) of the colons. We unblinded these five studies and determined the sensitivity and the total number of detections. On the basis of these results, which showed too many false-positive detections, we added two more filters (filters 6 and 7), which placed an upper limit of 1 on the sphericity.
Two more filters (filters 8 and 9) were added as points of reference (Table 2). Filter 8 shows the effect of selection criteria that are too broad, and filter 9 is for comparison with results in a prior publication (11). Filter 8 is similar to the primary criterion alone ("elliptic curvature of the peak subtype"), except that it also uses size and curvature criteria to discard tiny detections and detections likely to represent noise. Filter 9 is the same as the criteria in reference 11 and similar to filter 6. Its purpose is to show equivalency of the two size criteria (vertex count and maximum dimension).
All 20 studies were then analyzed by using all nine filters. For each study in which each filter was used (180 evaluations), a detection was considered true-positive when its location and size matched that of a polyp in the colonoscopy report. To make this assessment, a radiologist (R.M.S.) evaluated transverse source images, endoluminal three-dimensional images, and when needed, reformatted coronal and sagittal images. If more than one polyp was present in a segment, we used size to differentiate them to the best of our abilities. A polyp was considered to be detected when the computer algorithm was used to identify one or more vertices on its surface.
The sensitivity and total number of detections were determined for all 20 studies by using all nine filters. On the basis of these results, the optimum filter was identified. The optimum filter, as compared with the other filters, had both a high sensitivity and a low mean number of total detections per colon. Specificity was determined for this optimal filter on the basis of 160 colonic segments (eight segments per colon). The number of false-positive detections was determined by subtracting the number of detections of polyps from the total number of detections. There could be more than one detection of a polyp, especially if the polyp was large; that is, independent detections could occur in different parts of the polyp. The same polyp could appear twice in the data set if it was located in the 3 cm of overlap. In such cases, the polyp was considered to be detected if it was detected in either or both clusters of overlapping CT colonographic images.
Polyp Densitometry
For the optimal filter, a method for improving specificity was applied. With this method, the CT attenuation values of each voxel within a potential polyp are sampled along a single ray directed through the polyp. The purpose of this method is to identify voxels that have soft-tissue attenuation within the potential polyp. If any voxel along the ray exceeds a threshold, the detection is retained; conversely, if all voxels along the ray are lower than the threshold, the detection is discarded. Because of volume averaging, a threshold lower than 0 HU must be selected; empirically, a threshold of -124 HU removed a number of false-positive findings and no true-positive findings for the optimal filter. This method was also designed to avoid sampling voxels near the interface between the polyp and the air-filled lumen, at which volume averaging is maximal. The method was applied to only detections less than 1 cm. All detections 1 cm or greater were considered to be of potential clinical importance and retained.
The sampling ray was 5 mm long, began at the centroid of
the detected vertices that were along the edge of the polyp,
and pointed away from the colonic lumen in the direction of the
mean normal vector, which was determined by averaging the normal
vectors of the detected vertices. The equation for the sampling ray is
= 
- k
, where
is the location of the sampled voxel,
is the centroid,
k ranges from 0.5 to 5.0 mm in 0.5-mm increments, and
is the mean normal
vector. Voxels corresponding to the polyp surface were not sampled.
Colonic Distention and Segments
To evaluate the influence of adequate colonic distention on polyp detection, colons were given a distention score. Scoring was performed by a single radiologist (R.M.S.), who inspected a posteroanterior projection of the surface-rendered colon. A score of 0 indicated that the colon was collapsed; 1, partially distended; or 2, well distended. A score was given for each colonic segment on the basis of the amount of distention that predominated in that segment. Differentiation between partially and well-distended segments was based on expected colonic diameters for each segment in accordance with the observer’s experience with performing air-contrast barium enema examinations and from expected anatomic size distribution (ie, the cecum can distend more than the sigmoid colon).
The colon was divided into eight segments. Polyps were located in segments on the basis of the colonoscopy report, and colonic distention was scored on the basis of segments, as described previously. The segments were the rectum, sigmoid colon, descending colon, splenic flexure, transverse colon, hepatic flexure, ascending colon, and cecum. The location of one polyp was specified as a distance rather than as a segment in the colonoscopy report (eg, "Sixty-five centimeters from the anal verge"). For this polyp, a best location was determined by using computer graphics software to trace the distance along the colonic wall from the anal verge to the approximate location of the polyp (14). In one case, the colonoscopy report listed the polyp as being located in the hepatic flexure; however, in retrospect, it was clearly in the distal transverse colon. This was thought to represent a reporting error in a highly redundant colon.
Statistical Analysis
The Cochran Q test was performed to compare the sensitivity of the filters (17). Unpaired Student t tests were performed to compare polyp size distributions. P values of .05 or less were considered to indicate a significant difference.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Forty (80%) of the 50 polyps reported at colonoscopy were located on the CT scans and CT colonographic reconstructions when reference was made to the colonoscopy report. These included 26 of 28 large polyps, seven of 12 medium polyps, and seven of 10 small polyps. The remaining 10 could not be found and were considered to be false-negative findings.
The best colonic distention (mean score, 1.6–1.7) was achieved in the cecum and ascending colon. The poorest distention was in the sigmoid and descending colon and splenic flexure (mean score, 0.9–1.1). A level of distention intermediate to these was found in the rectum, hepatic flexure, and transverse colon (mean score, 1.5). Of 160 colonic segments, 74 (46%) were well distended, 64 (40%) were partially distended, and 22 (14%) were collapsed. A shaded surface display of a particularly well-distended colon is shown in Figure 1. There were 21 polyps (14 of which were large) in well-distended segments of colon, 28 (14 of which were large) in partially distended segments, and only one (not large) in a collapsed segment. There was no significant difference in mean size between polyps in well-distended segments, as compared with those in the remaining segments, either for polyps of all sizes (1.5 cm ± 1.4, n = 21 vs 0.9 cm ± 0.7, n = 29, respectively; P = .14) or for large polyps only (1.9 cm ± 1.5, n = 14 vs 1.5 cm ± 0.6, n = 14, respectively; P = .37).
|
|
|
|
|
|
The optimal filter, defined as having the highest sensitivity and fewer than 10 detections per colon, was filter 6, which had a mean of eight detections per patient and sensitivities of 40% (20 of 50) for all polyps and 64% (18 of 28) for large polyps. In well-distended colonic segments, the sensitivity of the optimal filter increased to 52% (11 of 21) for polyps of all sizes and 71% (10 of 14) for large polyps. At comparison, the sensitivity of the optimal filter for polyp detection in partially distended colonic segments was lower: 32% (nine of 28) for polyps of all sizes and 57% (eight of 14) for large polyps.
Results of tests of significance confirmed that the sensitivities for the primary shape filters (filters 1–7) were different. For polyps of all sizes, the Cochran Q statistic was 55.8 (P < .001) and for large polyps was 47.2 (P < .001).
There were 113 false-positive detections for filter 6 for the 20 colons, or a mean of six false-positive detections per colon. On the basis of colonic segments, there were 65 true-negative segments and 48 false-positive segments (specificity, 58%). When the ray sampling method was used (Fig 5), 44 false-positive detections were eliminated (39% improvement), without any reduction in sensitivity; this left 69 false-positive detections, or a mean of 3.5 false-positive detections per colon (78 true-negative segments and 35 false-positive segments; specificity, 69%).
|
The sensitivities for the lower isosurface thresholds were the same or slightly lower for the nine filters for all four groupings of polyps (all polyps, large polyps only, all polyps in well-distended segments, and large polyps in well-distended segments) in all but eight of 36 comparisons, without any obvious trend (Table 3). The mean number of detections per patient for each filter was comparable with that of the baseline thresholds.
The sensitivities for the higher edge-enhancing thresholds were mostly comparable with those of the baseline thresholds, with the exception of filters 1 and 5, with which sensitivity increases of 6%–36% were observed (Table 3). The largest increase occurred for filter 1 when large polyps in well-distended colonic segments were considered; with this filter, sensitivity increased from 21% to 57%. However, the mean number of detections per patient for each filter increased, as compared with that of the baseline thresholds for seven of the nine filters (all but filters 2 and 4).
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Although one might have expected real polyps to be more variable in shape and thus more difficult to detect than synthetic polyps, which were 10 mm in size and hemispheric in shape, we found comparable sensitivity for real polyps of sizes comparable with those of the synthetic ones. We demonstrated that a shape-based computer algorithm was used to detect real polyps at CT colonography with the patient supine with a sensitivity of 71% for large polyps in well-distended colonic segments; sensitivities were lower for smaller polyps and polyps in less well-distended colonic segments.
Our evaluation of a series of shape filters revealed the tradeoff in sensitivity and specificity as curvature thresholds, sizes, and sphericity varied. We also showed the limits of topologic analysis and the potential utility of techniques used to look inside the polyp and not just at its surface.
Sensitivity varied with the choice of shape filter. Curvature, size, sphericity, lumen threshold, and lesion attenuation were important parameters. Filter 8, which allowed the greatest range of acceptable curvature, size, and sphericity, was used to detect almost all (37 of 40) of the polyps that could be located at CT. However, there were too many false-positive results.
Comparison of filters 1 and 3 showed that sensitivity improved when maximum curvature values closer to 0 were used. Filter 8 showed that a minimum curvature of -4.0 per centimeter could be used to detect almost all of the polyps. Therefore, a minimum curvature of -4.0 to -3.0 and a maximum curvature of -1.0 to -0.5 per centimeter appear to be suitable for polyp detection.
The minimum size of a detected lesion also is important. Comparison of filters 2, 4, 5, and 8 showed that a minimum size of 120 vertices was too large but that a size of 2 vertices was too small. A minimum size of approximately 10–20 vertices appears to be suitable for polyp detection. Comparison of filters 6 and 9 showed that a size of 20 vertices, an estimate of areal measurement, was approximately equivalent in sensitivity to a linear measurement of 0.5 cm. Since the number of vertices depends on the resolution of the data, it is probably preferable to specify size criteria in centimeters rather than in number of vertices.
The sphericity criteria were added on the basis of our observation that false-positive detections tended to be curved in one direction and flat in the orthogonal direction; that is, they were close to having cylindric curvature. The sphericity dramatically decreased the number of detections, albeit at the cost of a slight decrease in sensitivity. Comparison of filters 3 and 6 and of filters 5 and 7 showed a 90% decrease in the number of detections and a decrease in sensitivity of 11–14% for large polyps. A sphericity of 1 or perhaps slightly greater will probably prove most suitable for polyp detection.
Experiments with three lumen thresholds, which determine the location of the interface between the lumen and the wall of the colon, showed that a nominal threshold (Table 1) was most appropriate. Results obtained with two other approaches thought to enhance the polyps by using either low or high thresholds were disappointing.
Lesion attenuation was another important parameter. Thirty-nine percent of the false-positive detections were eliminated by using a measurement of the lesion attenuation along a line that projected through the lesion. The sensitivity was unaffected.
The human observers had sensitivity 15%–24% higher than did the computer algorithm; this finding depended on the polyp size and the degree of colonic distention. However, the computer algorithm was used to detect one or more polyps missed by the human observers, who had a number of advantages. They were gastrointestinal radiologists with specialized training and made use of a variety of cues to discard false-positive results and increase true-positive results. For example, they used the entire CT image, not just a representation of its surface, to locate the presence of a homogeneous polypoid mass, identify the presence of air within stool that mimicked a polyp, identify an associated abnormality in the pericolonic fat, or determine that a bulge in the wall of the colon was due to extrinsic compression by another bowel loop or organ. In contrast, the computer algorithm relied heavily on the appearance of the colonic surface; when the surface was obscured (eg, in a poorly distended colonic segment), the computer algorithm was ineffective. The additional cues used by the human observers will need to be incorporated into the computer-assisted diagnosis algorithm to improve sensitivity.
The excess of false-positive findings (low specificity) is an important but not insurmountable problem. Many of the false-positive detections were plausible polypoid abnormalities and should have been reviewed by the radiologist anyway. Some of the false-positive detections were on the ileocecal valve, and others were readily recognized by a trained observer and discarded. There were 69 false-positive detections with ray sampling for the best filter (filter 6) in the 20 colons. For comparison, in a study of 70 colons (18), two observers identified 59 and 47 false-positive lesions that were 5 mm in diameter or greater. On the basis of these preliminary data, to achieve performance comparable with that of human observers, the number of false-positive detections needs to be reduced by a factor of about 5.
There are various approaches to reducing the number of false-positive detections. For example, the shape filters could be further optimized by using statistical discriminant analyses or neural networks primed with curvature data of known polyps and known false-positive detections. By reducing the number of false-positive detections, it might be feasible to use a more sensitive filter, combine filters, or use filters tailored to identify more small or medium polyps.
As we have seen, topologic analysis alone is probably insufficient to achieve the sensitivity and specificity needed for routine clinical use. Measurement of polyp attenuation is likely to be an important component of computer-assisted polyp detection. Although there are many possible ways to implement such a measurement, an advantage of the ray sampling method we described is that it does not require identification of the outer wall of the colon, which can be technically difficult to locate.
Although the sensitivity of our algorithm may improve with further refinement, it is not clear that it can exceed that of consensus reading by multiple trained radiologists; that is, the inherent sensitivity of CT colonography as currently practiced may be 75%–90% (7,8, 18,19). However, the automated polyp detector would not be subject to fatigue, and its results would be consistent and reproducible.
In the current study, 46% of the colonic segments were well-distended and 40% were partially distended. Of the major clinical CT colonographic studies in the literature, a few (7–9,20) provided a binary description of colonic distention (ie, adequately or inadequately distended), and only one (21) provided a semiquantitative measure of colonic distention that was similar to ours, which was based on visual assessment. Morrin et al (21) reported adequate distention of 185 of 200 colonic segments (93%, computed from supine scout image data presented therein); this was similar to our value of 86% (well-distended plus partially distended segments) even though they used a five-point scale, in contrast with our three-point scale. In our experience, the degree of colonic distention in the subjects in our study was typical of what can be expected in a general patient population.
We scanned the colon by using overlapping clusters of CT sections because our selected CT scanner protocol prevented scanning of the entire colon in a single breath hold. Overlapping clusters are not ideal because they can make interpretation more difficult. Multi–detector row CT scanners may eliminate the need for overlapping sections, since they can scan the entire colon in a single breath hold.
The current study had several limitations, some of which have been described previously (11). The patient population was biased, since all patients had polyps. This population was deemed suitable for a feasibility study. In addition, our data set was dominated by large polyps; this was an atypical size distribution (22). Large polyps are generally considered of greater clinical importance, however (23). Our algorithm will need to be validated with a larger, more representative sample to include patients without polyps. A measure of bowel preparation adequacy would be of value.
The sensitivity of computer-aided polyp detection with both prone and supine imaging must also be studied. Consensus is building that both prone and supine scanning are necessary to ensure adequate distention and displace fluid, which can obscure polyps on the dependent surface. Differences in colonic distention between the two types of scanning could enhance detection. Prone scanning could greatly improve sensitivity because a large fraction (17 [34%] of 50) of the polyps were in the sigmoid colon, which was among the least well-distended segments. The sensitivity for polyp detection with the human observers and the computer algorithm would likely be greater if prone CT colonographic data were added.
We did not measure the time required to produce an interpretation by using computer-aided diagnosis; therefore, the potential time savings promised by using this method needs to be determined. In addition, the effect on the cost-effectiveness of work-ups of false-positive polyps will need to be considered. Poor specificity could reduce the cost-effectiveness of CT colonography relative to that of conventional colonoscopy (10).
In clinical practice, the algorithm described here could be used to improve physician efficiency. It requires minimal training and could be executed by a technologist. Its results could be integrated with displays in which volume rendering or standard CT images were used, such as those now in common use for interpreting CT colonographic examinations (24). Potential polyps detected with the computer can be cross-referenced to a diagram of the colon to help the physician locate them (15). Perspective three-dimensional endoluminal renderings of the potential polyps could then be precomputed to avoid delays and displayed (15). The algorithm can run on a desktop personal computer; thus, computer hardware costs would be low.
In clinical use, the automated polyp detector could serve as a second reader. It has been shown in the mammographic literature (25) that the sensitivity for detecting breast masses increases when two radiologists interpret the results of each examination. However, a second reader increases the cost of performing an examination. An automated polyp detector with sufficient sensitivity could be used to reduce harmful false-negative interpretations, without the added time and expense of an additional physician interpreter. In our study, filters 6 and 7 each were used to identify two polyps that were missed by a human observer.
In conclusion, computer-assisted polyp detection with CT colonography is feasible. When further optimized, computer-assisted polyp detection may lead to greater use of CT colonography and an improvement in its performance.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
R.M.S., L.M.P., and J.D.M. have a patent pending, and C.D.J. and J.E.R. have received patents on the subject matter discussed in this article
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  N. Petrick, M. Haider, R. M. Summers, S. C. Yeshwant, L. Brown, E. M. Iuliano, A. Louie, J. R. Choi, and P. J. Pickhardt CT Colonography with Computer-aided Detection as a Second Reader: Observer Performance Study Radiology, December 1, 2007; 246(1): 148 - 156. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. G. Fletcher, F. Booya, R. M. Summers, D. Roy, L. Guendel, B. Schmidt, C. H. McCollough, and J. L. Fidler Comparative Performance of Two Polyp Detection Systems on CT Colonography Am. J. Roentgenol., August 1, 2007; 189(2): 277 - 282. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. Jensch, R. E. van Gelder, J. Florie, M. A. T.-d. Graaf, J. V. Lobe, P. M. M. Bossuyt, S. Bipat, C. Y. Nio, and J. Stoker Performance of Radiographers in the Evaluation of CT Colonographic Images Am. J. Roentgenol., March 1, 2007; 188(3): W249 - W255. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. D. O'Connor, R. M. Summers, J. Yao, P. J. Pickhardt, and J. R. Choi CT Colonography with Computer-aided Polyp Detection: Volume and Attenuation Thresholds to Reduce False-Positive Findings Owing to the Ileocecal Valve Radiology, November 1, 2006; 241(2): 426 - 432. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. Shi, P. Schraedley-Desmond, S. Napel, E. W. Olcott, R. B. Jeffrey Jr, J. Yee, M. E. Zalis, D. Margolis, D. S. Paik, A. J. Sherbondy, et al. CT Colonography: Influence of 3D Viewing and Polyp Candidate Features on Interpretation with Computer-aided Detection. Radiology, June 1, 2006; 239(3): 768 - 776. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. A. Taylor, S. Halligan, A. Slater, V. Goh, D. N. Burling, M. E. Roddie, L. Honeyfield, J. McQuillan, H. Amin, and J. Dehmeshki Polyp Detection with CT Colonography: Primary 3D Endoluminal Analysis versus Primary 2D Transverse Analysis with Computer-assisted Reader Software Radiology, June 1, 2006; 239(3): 759 - 767. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. A. Taylor, S. Halligan, D. Burling, M. E. Roddie, L. Honeyfield, J. McQuillan, H. Amin, and J. Dehmeshki Computer-Assisted Reader Software Versus Expert Reviewers for Polyp Detection on CT Colonography. Am. J. Roentgenol., March 1, 2006; 186(3): 696 - 702. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. T. Johnson, C. D. Johnson, J. G. Fletcher, R. L. MacCarty, and R. L. Summers CT Colonography Using 360{degrees} Virtual Dissection: A Feasibility Study Am. J. Roentgenol., January 1, 2006; 186(1): 90 - 95. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. KOBASHI, K. KONDO, and Y. HATA Computer-Aided Diagnosis of Intracranial Aneurysms in MRA Images with Case-Based Reasoning IEICE Trans D: Information, January 1, 2006; E89-D(1): 340 - 350. [Abstract] [PDF]
|
|
|
|
|
|
M. Macari and E. J. Bini CT Colonography: Where Have We Been and Where Are We Going? Radiology, December 1, 2005; 237(3): 819 - 833. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
C. L. Partain, H.-P. Chan, J. G. Gelovani, M. L. Giger, J. A. Izatt, F. A. Jolesz, K. Kandarpa, K. C. P. Li, M. McNitt-Gray, S. Napel, et al. Biomedical Imaging Research Opportunities Workshop II: Report and Recommendations Radiology, August 1, 2005; 236(2): 389 - 403. [Full Text] [PDF]
|
|
|
|
|
|
G. Iinuma, N. Moriyama, M. Satake, K. Miyakawa, U. Tateishi, N. Uchiyama, T. Akasu, T. Fujii, and T. Kobayashi Vascular Virtual Endoluminal Visualization of Invasive Colorectal Cancer on MDCT Colonography Am. J. Roentgenol., April 1, 2005; 184(4): 1194 - 1198. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. M. Summers, M. Franaszek, M. T. Miller, P. J. Pickhardt, J. R. Choi, and W. R. Schindler Computer-Aided Detection of Polyps on Oral Contrast-Enhanced CT Colonography Am. J. Roentgenol., January 1, 2005; 184(1): 105 - 108. [Full Text] [PDF]
|
|
|
|
 |
|
 K Doi Current status and future potential of computer-aided diagnosis in medical imaging Br. J. Radiol., January 1, 2005; 78(suppl_1): S3 - s19. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. M. Summers, J. Yao, and C. D. Johnson CT Colonography with Computer-aided Detection: Automated Recognition of Ileocecal Valve to Reduce Number of False-Positive Detections Radiology, October 1, 2004; 233(1): 266 - 272. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 P. B. Cotton, V. L. Durkalski, B. C. Pineau, Y. Y. Palesch, P. D. Mauldin, B. Hoffman, D. J. Vining, W. C. Small, J. Affronti, D. Rex, et al. Computed Tomographic Colonography (Virtual Colonoscopy): A Multicenter Comparison With Standard Colonoscopy for Detection of Colorectal Neoplasia JAMA, April 14, 2004; 291(14): 1713 - 1719. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
T. M. Gluecker, J. G. Fletcher, T. J. Welch, R. L. MacCarty, W. S. Harmsen, J. R. Harrington, D. Ilstrup, L. A. Wilson, K. E. Corcoran, and C. D. Johnson Characterization of Lesions Missed on Interpretation of CT Colonography Using a 2D Search Method Am. J. Roentgenol., Ap |
