Segmentation, shape, and volume analysis improve sensitivity

CAD techniques refine CT colonography

By: Hiroyuki Yoshida, Ph.D., And Abraham Dachman, M.D

Colorectal cancer is the second most common cause of cancer deaths in the U.S., with approximately 60,000 deaths per year.1 The compliance of the population with current screening strategies is poor, and the development of a new structural examination of the entire colon has attracted the attention of patients and physicians alike. CT colonography, also known as virtual colonoscopy, has emerged as a viable alternative for examining the colon for polyps and masses and as a potential screening tool.2 The test typically involves supine and prone CT scans of the cleansed colon, after insufflation with air or carbon dioxide to a level that the patient can tolerate, and interpretation with specialized software of both 2D and endoluminal 3D images.

Advances in image software allow radiologists to interpret CT colonography by various means, but the large amount of data generated and the steep learning curve for interpretation create problems. Accumulating evidence supports the promise of high-detection performance for polyps in CT colonography,3 but three difficulties must be addressed: Interpretation of an entire CT colonography examination is time-consuming (10 to 40 minutes even for experts), the interpretation is prone to perceptual errors, and the diagnostic performance is variable among readers.2,4,5

Computer-aided detection of polyps has been generating considerable excitement. CAD is attractive because it has the potential to overcome these difficulties in CT colonography.6-11 A CAD system automatically detects polyps in CT images and provides the locations of suspicious ones to radiologists. The "second opinion" offered by CAD has the potential to reduce radiologists' interpretation time because they can focus on the small number of regions indicated as suspicious by CAD and quickly survey a large portion of the colon that is likely to be normal. Like CAD for mammography, high-performance CAD in CT colonography has the potential to reduce perceptual errors and variability among readers, improving detection performance.12,13

Studies suggest that, for CT colonography to be cost-effective for colorectal polyp screening, the interpretation time needs to be reduced substantially to the range of three to five minutes while a high sensitivity of more than 80% is maintained.14 CAD can be a key technology in moving CT colonography from the research arena to routine clinical practice and, especially, to the screening setting.14 In the past several years, researchers have developed prototype CAD systems for the automated detection of polyps with CT colonography.6,7,9

AUTOMATED POLYP DETECTION

Most CAD schemes consist of the following steps: extraction of the colon from 3D volume generated from axial CT colonographic images, detection of polyp candidates in the extracted colon, removal of false positives from the polyp candidates, and display of detected polyps. Key techniques for the first three steps are described below.

- Extraction of the colon. Several fully automated methods for extraction of the colon have been developed.15,16 One of the most effective approaches is the extraction of a thick region that encompasses the entire colon.17 Although several investigators use a surface-generation method for extraction of the inner surface of the colonic wall,15,18,19 this risks losing a part of a polyp, in particular, the internal structure. The approach based on a thick region can extract a major portion of a polyp, including its internal structure, and thus can detect polyps more accurately than is possible by the surface-generation method.

A group at the University of Chicago has developed a two-step method of knowledge-guided segmentation20 for fully automated extraction of the thick region encompassing the entire colon. The first step removes the normal structures that are not connected to the colon, such as the area surrounding the body region ("outer air"), the osseous structures (spine, pelvis, and parts of the ribs), and the lung bases, based on thresholding of the 3D volume with the CT values characteristic of these structures. The colon is then segmented from the remaining region by thresholding of the CT and gradient magnitude values corresponding to the colonic wall. The resulting segmentation contains extracolonic structures, such as the stomach and small bowel, that are connected to the colon. The second step applies a self-adjusting volume-growing method to the colonic lumen surrounded by the colonic wall identified in the first step. Because most parts of the small bowel and stomach adhering to the colon are in contact with the colonic wall, but not with the colonic lumen, intersection of this grown region and the colonic wall removes these adhering extracolonic structures and thus determines the final region encompassing only the colon. A further advantage is that this approach can reliably extract a partly collapsed colon. Figure 1 shows an example of the knowledge-guided segmentation process.

-Detection of polyp candidates. After the entire colon is extracted, polyp candidates are detected by extraction of geometric features that characterize polyps at each point in the colonic wall. The colonic wall appears as a large, nearly flat, cuplike structure (Figure 2). Methods for detection of polyps need to characterize these shape differences among polyps, folds, and the colonic wall. Various methods address this need, including volumetric shape index,7 surface curvature,19 surface normal overlap,21 and sphere fitting.11

The volumetric shape index is shown to be highly effective in characterizing these shapes. The shape index analyzes the vicinity of a voxel and places a voxel in one of five topologic classes: cup, rut, saddle, ridge, or cap. The highest shape index values that correspond to a caplike shape are of particular interest because many polyps exhibit this shape. Color coding of the anatomic structures in the colonic lumen based on the shape index can differentiate among polyps, folds, and colonic walls (Figure 3).

After calculation of the geometric feature, the polyp candidates are detected by hysteresis thresholding, which extracts sets of spatially connected voxels having shape index and curvedness values characteristic of polyps, followed by conditional morphological dilation,22 which extracts the entire region corresponding to a polyp by iteratively adding a layer of voxels to the surface of the region extracted by hysteresis thresholding.

-Removal of false-positive detections. The polyp candidates thus detected may include false positives, and methods for reducing their number include CT attenuation,6 random orthogonal shape section,10 optical flow, and texture analysis.24

Studies have shown that prominent folds and stool are major sources of false positives.8,9 A gradient-based feature called gradient concentration is particularly effective for reducing false positives due to folds.24 This feature measures the degree of concentration of the gradient orientations of the CT values in the vicinity of a voxel. Generally, polyps are hemispherical objects on the colonic wall, and the gradient vectors at voxels on the boundary between the polyp and luminal air converge to a point deep in the center of the polyp. Folds, on the other hand, are elongated and ridgelike. The gradient vectors at voxels on the boundary between the fold and luminal air converge to a line. Because of these differences, the distribution of the gradient concentration at a voxel can be an effective means for differentiating folds from polyps.

Stool differentiation is based on differences in the internal density variation between polyps and stool. These density variations are caused by the tendency of stool to contain air bubbles that can be recognized on CT images as an inhomogeneous textural pattern, or mottle pattern. In contrast, polyps tend to have a homogeneous textural pattern, or solid pattern, without intratumoral air. Thus, the use of the variance of CT values that characterize the homogeneity of the CT density within a polyp is helpful.24

The final detected polyps are obtained by application of a statistical classifier, which is based on the image features extracted in these steps, to the differentiation of polyps from false positives. Researchers use nonparametric classifiers such as neural networks11 and a support vector machine.10 Any combination of features and a classifier that provides high classification performance, however, should be sufficient for the differentiation task. The group at the University of Chicago developed feature-guided analysis22 for determination of a representative subregion of a polyp for calculation of the above two textural features and then applied quadratic discriminant analysis to the feature space generated by the shape index and the textural features. Quadratic discriminant analysis is a parametric classifier that generates a decision boundary that optimally partitions the feature space into a polyp class and a false-positive class based on supervised learning.

PROMISE AND PITFALLS

Several academic institutions are conducting early clinical trials to demonstrate the performance of their CAD schemes. They compare the locations of the polyps detected by CAD with the true locations of polyps that are determined visually in CT colonography data sets. In a by-patient analysis, the CAD scheme developed at the University of Chicago yielded 100% sensitivity with 1.3 false positives per patient. In a by-polyp analysis, the sensitivity was 95% (20 out of 21), with 1.5 false positives.

These preliminary results indicate that CAD in detection of polyps is promising, with high sensitivity and a low false-positive rate. Extension to a larger database, retrospectively and prospectively, will be needed for confirmation of the performance of these CAD schemes.

The types of pitfalls included in the CAD results are similar to those experienced by radiologists.8,9 Most of the CAD schemes assume that a polyp appears to be a polypoid lesion. Sessile polyps that do not protrude sufficiently into the lumen, those distorted by normal structures, and those that lose their parts due to the partial volume effect may be missed by CAD. Figure 4 shows an 8-mm polyp missed by our CAD scheme. The polyp was located at a narrow valley where two folds merged, and thus the shape of the polyp was distorted from a polypoid shape.

Studies conducted at the University of Chicago show the major causes of false positives detected by the CAD scheme:7,25 Approximately 40% were caused by folds (Figure 5). These consisted of sharp folds at the sigmoid, folds prominent on the colon wall, two converging folds, ends of folds in the tortuous colon, and folds in a colon that was not sufficiently distended. About 20% were caused by retained stool (Figure 6), which is often a major source of error for radiologists. Around 15% were caused by residual materials inside the small bowel and stomach, and 10% were caused by ileocecal valves.

All of these false positives exhibited polyplike shapes,8,25 but most could easily be distinguished from polyps by experienced radiologists. False positives due to stool, for example, are distinguishable from polyps by comparison of both prone and supine views and by examination of the cut-plane view to visualize the mottle pattern characteristic of stool. False positives due to ileocecal valves can be differentiated from polyps based on the characteristic location and appearance of the valve.

FUTURE CHALLENGES

CAD for CT colonography is still under investigation, and a number of challenges must be met.26 Although the preliminary results are promising, the detection performance of CAD needs to improve, along with the determination of the optimal CT colonography protocol for CAD, including CT parameter settings. The resulting CAD should be evaluated based on multicenter, prospective clinical trials, rather than on a single-center, retrospective study. The ultimate goal of CAD is to improve the diagnostic performance of radiologists in the detection of polyps. Therefore, observer performance studies should be conducted for evaluation of the diagnostic performance of a radiologist with and without computer aid in a prospective fashion. Such studies have been conducted for CAD for mammography12 and chest radiography,13 successfully demonstrating the benefit of CAD.

Considerable interest has accompanied the application of stool subtraction techniques15,27 and low-dose CT scanning.28 Stool tagging changes the CT values of liquid stool, and thus it may improve the performance of CAD in differentiating polyps from stool. New CAD algorithms need to be developed, and fecal tagging will provide new challenges for CAD. Low-dose CT colonography is especially important for colon cancer screening, but low-dose CT scans come at the cost of increased image noise. Therefore, we need to investigate the effect of noise due to low dose on the performance of CAD. No research has yet been performed on this problem.

CAD for CT colonography is still under investigation. It shows potential for detecting polyps with high sensitivity and with a clinically acceptable low false-positive rate. If the preliminary results of CAD can be translated into a larger number of cases, radiologists will need to interpret only the few regions indicated by the CAD scheme in a CT colonography examination. Such an aid is expected to reduce the interpretation time substantially while maintaining or improving the diagnostic accuracy. CAD has the potential to bring CT colonography one step closer to cost-effective clinical practice, and especially to the screening setting.

DR. YOSHIDA is an assistant professor of radiology, and DR. DACHMAN is a clinical professor of radiology, both at the University of Chicago.

---

References

1. Potter JD, Slattery ML, Bostick RM, Gapstur SM. Colon cancer: a review of the epidemiology. Epidemiol Rev 1993; 15:499-545.
2. Johnson CD, Dachman AH. CT colonography: the next colon screening examination? Radiology 2000;216:331-341.
3. Dachman AH. Diagnostic performance of virtual colonoscopy. Abdominal Imaging 2002;27:260-267.
4. Pescatore P, Glucker T, Delarive J, et al. Diagnostic accuracy and interobserver agreement of CT colonography (virtual colonoscopy). Gut 2000;47:126-130.
5. Fletcher JG, Johnson CD, Welch TJ, et al. Optimization of CT colonography technique: prospective trial in 180 patients. Radiology 2000;216:704-711.
6. Summers RM, Johnson CD, Pusanik LM, et al. Automated polyp detection at CT colonography: feasibility assessment in a human population. Radiology 2001;219:51-59.
7. Yoshida H, Nappi J. Three-dimensional computer-aided diagnosis scheme for detection of colonic polyps. IEEE Transactions on Medial Imaging 2001;20:1261-1274.
8. Yoshida H, Nappi J, MacEneaney P, et al. Computer-aided diagnosis scheme for the detection of polyps in CT colonography. RadioGraphics 2002;22:963-979.
9. Yoshida H, Masutani Y, MacEneaney P, et al. Computerized detection of colonic polyps at CT colonography on the basis of volumetric features: pilot study. Radiology 2002;222:327-336.
10. Gokturk SB, Tomasi C, Acar B, et al. A statistical 3-D pattern processing method for computer-aided detection of polyps in CT colonography. IEEE Trans Med Imaging 2001;20:1251-1260.
11. Kiss G, Van Cleynenbreugel J, Thomeer M, et al. Computer-aided diagnosis in virtual colonography via combination of surface normal and sphere fitting methods. Eur Radiol 2002;12:77-81.
12. Chan HP, Sahiner B, Helvie MA, et al. Improvement of radiologists' characterization of mammographic masses by using computer-aided diagnosis: an ROC study. Radiology 1999;212:817-827.
13. Kobayashi T, Xu XW, MacMahon H, et al. Effect of a computer-aided diagnosis scheme on radiologists' performance in detection of lung nodules on radiographs. Radiology 1996;199:843-848.
14. Sonnenberg A, Delco F, Inadomi JM. Cost-effectiveness of colonoscopy in screening for colorectal cancer. Ann Intern Med 2000;133:573-584.
15. Chen D, Liang Z, Wax MR, et al. A novel approach to extract colon lumen from CT images for virtual colonoscopy. IEEE Trans Med Imaging 2000;19:1220-1226.
16. Wyatt CL, Ge Y, Vining DJ. Automatic segmentation of the colon for virtual colonoscopy. Comput Med Imaging Graph 2000;24:1-9.
17. Masutani Y, Yoshida H, MacEneaney P, Dachman A. Automated segmentation of colonic walls for computerized detection of polyps in CT colonography. JCAT 2001;25:629-638.
18. Hong L, Liang Z, Viswambharan A, et al. Reconstruction and visualization of 3D models of colonic surface. IEEE Transactions on Nuclear Science 1997; 44:1297-1302.
19. Summers RM, Beaulieu CF, Pusanik LM, et al. Automated polyp detector for CT colonography: feasibility study. Radiology 2000;216:284-290.
20. Nappi J, Dachman AH, MacEneaney P, Yoshida H. Automated knowledge-guided segmentation of colonic walls for computerized detection of polyps in CT colonography. JCAT 2002;26:493-504.
21. Paik DS, Beaulieu CF, Jeffrey RB, et al. Computer aided detection of polyps in CT colonography: method and free-response ROC evaluation of performance. Radiology 2000;217(P):370.
22. Nappi J, Yoshida H. Feature-guided analysis for reduction of false positives in CAD of polyps for CT colonography. Medical Physics 2002:(submitted).
23. Acar B, Beaulieu CF, Paik DS, et al. Computer-aided detection of colonic polyps in CT colonography using optical flow. Radiology 2001;221(P):331.
24. Nappi J, Yoshida H. Automated detection of polyps with CT colonography: evaluation of volumetric features for reduction of false-positive findings. Acad Radiol 2002;9:386-397.
25. Yoshida H, Nappi J, MacEneaney P, et al. Computer-aided viasualization and differnetiation of stool from polyps in CT colonography. Radiology 2001;221(P):653.
26. Summers R, Yoshida H. Future directions of CT colonography: computer-aided diagnosis. In: Dachman AH, ed. Atlas of virtual colonoscopy. New York: Springer-Verlag, 2002.
27. Zalis ME, Hahn PF. Digital subtraction bowel cleansing in CT colonography. AJR 2001; 176:646-648.
28. Macari M, Bini EJ, Xue X, et al. Colorectal neoplasms: prospective comparison of thin-section low-dose multi-detector row CT colonography and conventional colonoscopy for detection. Radiology 2002;224:383-392.
29. Paik DS, Beaulieu CF, Mani A, et al. Evaluation of computer-aided detection in CT colonography: potential application to a screening population. Radiology 2001;221(P):332.

Close Window