Abdom Imaging 27:268 –274 (2002) DOI: 10.1007/s00261-001-0168-7

Abdominal Imaging © Springer-Verlag New York Inc. 2002

Challenges for computer-aided diagnosis for CT colonography R. M. Summers Diagnostic Radiology Department, National Institutes of Health, Building 10, Room 1C660, 10 Center Drive, MSC 1182, Bethesda, MD 208921182, USA

Abstract Computer-aided diagnosis for computed tomographic colonography is in its infancy but has the potential to improve sensitivity and decrease costs for colonic polyp detection. This article reviews the current state of research in this nascent field and explores major challenges and avenues for future work. Key words: Computed tomography, colon—Computed tomography, three-dimensional reconstruction—Colon cancer—Image processing.

Although the methods for performing and interpreting computed tomographic colonography (CTC) studies are evolving, two prominent problems have emerged that are yet to be fully addressed: a lengthy interpretation time and differences in sensitivity among different observers [1]. Computer-aided diagnosis (CAD) might address these problems. CAD is particularly challenging in CTC because the colon is highly deformable and there are multiple polyp mimics. In this article, I review some of the challenges facing CAD in the colon. I also review how CTC is currently interpreted, how one can apply CAD to CTC, and what lessons one can learn from other types of CAD.

The target lesion The objective of CTC is to identify polyps and cancers of the colon. Colonic polyps are small growths that can range from a millimeter to several centimeters. Polyps can be flat, have a broad base of attachment (sessile), or be at the end of a stalk of variable length (pedunculated). On CT, the attenuation of a polyp is similar to that of soft tissue (such as muscle and solid organs). Lipomatous (fat containing) polyps can be readily identified by their low central attenuation values. Cancers tend to be larger than

polyps (from which they commonly arise) and may appear as focal or circumferential masses.

Learning from the radiologist A radiologist uses multiple criteria for identifying polyps on CTC. Although algorithms could be developed by learning from positive and negative cases (training by example), algorithms are likely to be more successful if based on features that are proven clinically relevant (i.e., based on anatomy or pathology). Examples of features radiologists find useful include anatomic knowledge, CT attenuation, shape, and texture. Unlike current computer algorithms, radiologists know which portion of the colon they are examining. This anatomic knowledge may be important if CAD algorithms have different sensitivities in different parts of the colon (Shuo-Hung Ling, personal communication). In addition, anatomic localization can help to distinguish normal structures from pathology. For example, the ileocecal valve can mimic a polyp but the ileocecal valve is always found at a particular location within the cecum. Anatomic localization will be more difficult in a patient for whom portions of the colon have been resected. Radiologists can recognize a number of polyp mimics by using texture and shape. Image texture can help the radiologist distinguish retained fecal matter from polyps because stool might contain small bubbles of gas. Haustral folds (normal features located every few centimeters that separate the colon into a series of sacculations) often can be distinguished from polyps by using three-dimensional reconstructions that reveal their characteristic shape. In addition to texture and shape, other features used by radiologists are CT attenuation and thickness of the colonic wall. A polyp should have soft tissue density, and thickness of the colonic wall should be greater than that of the adjacent normal colon. The normal colonic wall typically has a thickness of 1 or 2 mm, depending on dis-

R. M. Summers: Challenges for CTC CAD

tention of the colon, with apparent wall thickness decreasing with greater distention. It is not known whether flat polyps would be missed with these wall thickness cutoff values.

Learning from other radiology CAD There is a large body of knowledge developed over the past 10 –15 years in the field of radiology CAD. Two major areas of radiology CAD are breast and lung cancer diagnoses. Just as clinical CTC researchers can learn from mammography [2], CTC CAD researchers can learn from mammography CAD research. For example, mammography CAD researchers have found that different types of lesions (masses, microcalcifications) require different approaches to detection. Mammography CAD researchers have explored a wide range of feature extraction and classification methodologies. They have also analyzed statistical aspects in great detail. Although there are many differences between the clinical problems in mammography and CTC, the prior work forms a foundation on which CTC CAD can be built and expanded. As in mammography CAD, specific diagnoses will require distinct approaches. For example, in CTC CAD, polyps and cancers may require different techniques. Polyps tend to be smaller and localized. Cancers may be larger and circumferential. In mammography CAD, some systems extract a large number of features (⬎25) for classification purposes. Examples of useful features are fractal dimension, spicularity, and spatial gray-level dependence matrices [3, 4]. CTC CAD currently extracts fewer features (⬍10), suggesting room for further development. The situation is similar for lesion segmentation, which is also further developed in mammography CAD than in CTC CAD. Mammography CAD researchers benefit from at least two publicly accessible databases containing hundreds of proven cases (none is yet available for CTC). Such databases are expensive to produce but are likely to prove crucial for algorithm research and development. Lung cancer CAD research has led to automated lung nodule detection on chest radiographs and CT scans. Just as in mammography CAD, image segmentation has been studied extensively. For example, the problems of lung nodules that abut the pleura, mediastinum, and pulmonary vessels have been investigated. Changes in nodule size over time can be computed automatically and new nodules (change analysis) can be detected [5]. False positive nodules have been studied extensively to improve specificity. As in mammography CAD, large image databases have been created. These parallels between lung cancer and breast cancer CAD suggest that similar approaches will be required for CTC CAD.

269

Measurements useful for CTC CAD Analyses of surface topology and CT attenuation are important aspects of clinical diagnosis of colonic polyps on CTC. The National Institutes of Health (NIH) has implemented both in its CAD system [6, 7]. Preliminary reports from other research groups have indicated success with similar techniques [8 –11]. The current NIH CAD system performs topologic analysis by analyzing the local shape of the inner colonic wall to identify abnormalities protruding into the lumen (Figs. 1, 2) [6, 7]. Curvature analysis has been a successful means of implementing a topologic analysis. Once abnormalities of colonic shape are determined by the topologic analysis, “virtual biopsy” of the suspected abnormality is done by sampling the CT attenuation along rays through the colonic wall to determine whether a mass is present or the wall is thickened [7]. If intravenous contrast is given (not routinely done in current CTC), contrast enhancement can be detected at this stage. In my experience, analysis of topology and CT attenuation can be performed quite rapidly (⬍5 min) [7]. The NIH is in the process of studying textural analysis, another technique that may prove useful for CAD. Textural analysis consists of inspecting the CT image for locally homogeneous regions of soft tissue attenuation that might represent polyps. The search for such abnormalities is done within and in close proximity to the colonic wall. In the following sections, I describe 10 challenges that I believe must be addressed in a successful CTC CAD system. Some of the challenges have analogs in other medical imaging CAD, but some are unique. They occur at different stages of the analysis pipeline, from how to determine ground truth in an image database to how to integrate CAD effectively into clinical practice.

Challenge 1: The difficulty of specifying polyp location An important step in the development of a successful CAD algorithm is to create a well-annotated database of cases that represent “ground truth.” Ground truth consists of identifying patients with polyps and determining the precise location of each polyp. The precise location is important because classifiers can only be trained if they are given accurate information about the shape of the polyp and its relation to adjacent structures. At the NIH, we have found the task of matching polyps on conventional colonoscopy and CTC to be challenging. In practice, identifying the precise coordinates of a known polyp is difficult and can be a source of error. We believe this problem is difficult because many polyps do not fit the common notion of the spherical protuber-

270

R. M. Summers: Challenges for CTC CAD

Fig. 1. Detected 1-cm colonic polyp (arrows). A Three-dimensional surface rendering of colon. B Axial CTC image. C Color-coded threedimensional endoluminal surface reconstruction. The red color indicates CAD detection of the polyp.

ance. Polyps are often on folds, may be tubular in shape, on a long floppy stalk, or flat.

Challenge 2: Tradeoffs between sensitivity and specificity A second objective of CAD research in general is to achieve high sensitivity of detection with the fewest number of false positives. As with other types of CAD, there is a tradeoff between sensitivity and specificity. Sensitivity can be increased by using broader, more inclusive, classification criteria, but invariably this occurs at the expense of an increase in the number of false positives.

Because of the complexity of the normal colon and the diverse shapes and imaging characteristics of polyps and cancers, there is no simple relation between sensitivity and false positive detections. Tradeoffs between sensitivity and false positive detection can be determined only by evaluating various detection strategies on image databases proven by use of a gold standard (such as conventional colonoscopy). Image databases with proven and well-annotated cases are likely to prove crucial to these efforts. Free-response receiver-operating characteristics analyses are useful for portraying and evaluating these tradeoffs [10]. Free-response rather than standard receiver-operating characteristics analysis is needed to take lesion location and multiple lesions into account.

R. M. Summers: Challenges for CTC CAD

271 Fig. 2. A 1.5-cm rectosigmoid polyp detected by CAD. A Three-dimensional endoluminal reconstruction of air-insufflated colon on CTC showing polyp (large arrows). Red color indicates CAD detection. Several haustral folds are visible (small arrows). B, C Axial CT sections (B) before and (C) after labeling of the polyp (arrows) with CAD. C The labels are caps of white voxels along the tip of the polyp. Reproduced from Summers RM, Pusanik LM, Malley JD, et al. Method of labeling colonic polyps at CT colonography using computer-assisted detection. In: Computer assisted radiology and surgery (CARS). San Francisco: Elsevier Science, 2000:785–789; with permission from Elsevier Science.

Neural network algorithms and other types of artificial intelligence may be critical to determine the appropriate weights in the parameter space of various algorithms.

Challenge 3: Polyp camouflage An important objective of CTC interpretation is to distinguish true polyps from structures that mimic polyps and cancers. There are a number of polyp mimics, some of which have been discussed, such as stool, the ileocecal valve, and thick folds in underdistended segments. Also problematic are situations that hide polyps, and there is some overlap with the polyp mimics, including stool, underdistended segments, retained fluid, and the rectal tube. The computer-aided detector needs to identify and distinguish these findings that might camouflage polyps. For example, residual fluid can be recognized by its location in the dependent portion of the colon, the flat

air–fluid level, or its increased CT attenuation if the colonic fluid had been opacified. It may be possible to “subtract” residual fluid or stool from CTC images or mark them so they can be recognized [12, 13]. Methods for tagging stool using orally or rectally administered contrast agents may be useful for identifying and eliminating voxels containing stool. In addition, stool may be mobile and typically will appear in the dependent portion of the colon. Therefore, the prone and supine examinations will need to be compared.

Challenge 4: Using the supine and prone images together Evidence gathered in the past few years supports the notion that adding prone scanning to the original supine scan can improve diagnostic sensitivity [1]. The challenge is to determine how a computer algorithm can make use

272

R. M. Summers: Challenges for CTC CAD

Extrinsic markers such as loops of small bowel, blood vessels, visceral organs such as the liver, spleen, or kidneys, or other nonspecific densities in the fat adjacent to the colon could be used by a computer algorithm. Such registration is challenging because of the nonrigid nature of the transformation that occurs between supine and prone positioning. This is particularly true of the bowel, which hangs from a mesentery and various ligaments, and can assume markedly different configurations depending on gravity and pressure from adjacent structures. One possible approach to prone–supine registration is to open and flatten the colon [15–19] (much as a pathologist might do with an autopsy specimen) and then to register the folds. One problem we have encountered with such algorithms is that they tend to distort the size of potential abnormalities.

Challenge 5: Cancer detection

Fig. 3. Haustral fold detection for CTC. The manual count of haustral folds was 146, and the automated method counted 129 haustral folds (green), a few of which were false positive. Haustral fold identification might be useful for reconciling the supine and prone images, but this method is limited because folds are easily effaced by overdistention or obscured by underdistention or fluid.

of the prone and supine scans together rather than simply using them individually and separately to arrive at a diagnosis. For example, researchers have learned that other types of CAD benefit by matching suspected lesions identified on multiple views [14]. However, preliminary evidence suggests that polyps are often visible on only the prone or supine view but not on both. Radiologists use a variety of cues intrinsic and extrinsic to the colon to match the supine and prone segments. For example, the geometry of the haustral folds, the density and geometry of the potential abnormality, and relations with adjacent organs outside the colon may be important for proper registration of the images. A computer algorithm also can use features intrinsic and extrinsic to the colon. This falls into the image processing category of a nonrigid registration. One possible method is to construct a coordinate system based on the length of the colon. Another possibility is to enumerate haustral folds as intrinsic fiducials for registration or polyp localization (Fig. 3). This will be very challenging because overdistention of the colon can efface haustral folds and underdistention can result in collapsed segments in which it will be impossible to count folds.

Another important task in CTC is to detect cancers. Cancer detection poses challenges in addition to those of polyp detection because a polyp is a focal abnormality but a cancer can appear as a mass or a circumferential abnormality with different shape characteristics. Circumferential (napkin ring) cancers will cause predominantly lumen narrowing and wall thickening. Wall thickening and lumen narrowing need to be distinguished from peristalsis and underdistention. In general, peristalsis causes lumen narrowing without significant wall thickening and is not in the same location on prone and supine images. Colon cancers can sometimes produce abnormalities in the adjacent extracolonic fat. An automated lesion detector would need to evaluate the pericolonic fat for abnormality that could suggest tumor in the adjacent colon.

Challenge 6: Artifacts Artifacts are another type of polyp camouflage. A number of different artifacts can be recognized, including streak artifact from residual barium, surgical clips, and hip prostheses. Such artifacts will pose additional challenges for CAD. Threshold-based methods may be able to remove voxels containing residual barium. Streak artifact may be identifiable by its linear nature but in general will be challenging to eliminate.

Challenge 7: Poor colonic distention Adequate colonic distention is necessary for polyp detection by CTC. When the colon is inadequately distended or collapsed, the polyps are not outlined by air and their shape may be obscured. Fortunately, existing methods for colonic distention obtain partial or complete distention in

R. M. Summers: Challenges for CTC CAD

approximately 85–90% of colonic segments [7]. Sensitivity may be reduced in less well-distended segments [7]. In practice it is difficult to fully distend all colonic segments on supine and prone scans. For example, it is especially difficult to maintain distention in the transverse colon when the patient is prone because this portion of the colon is anterior and easily compressed by the patient’s weight. Placing a cushion under the lower chest may improve distention in the transverse colon. Other maneuvers during acquisition of the CT data may improve distention in other segments.

Challenge 8: Noisy images An objective of clinical CTC research is to detect polyps and cancers at the lowest possible radiation dose so as to reduce radiation exposure to the population. Low radiation dose is an especially important goal for colon cancer screening and is another reason low-dose CTC has been advocated [20]. Because of the large amount of image contrast between the air in the colon and the wall of the colon, low-dose scanning is feasible and typically produces excellent results. However, low-dose CTC images are noisier than higher-dose images. Increasing image noise may impair sensitivity and produce unacceptably noisy, three-dimensional reconstructions for virtual endoscopy. The sensitivity decreases because noise tends to distort the shape of polyps, thereby interfering with shape-based detection. In addition, preliminary evidence suggests that increasing image noise will also increase the number of false positive detections, thus impairing specificity. Therefore, image processing algorithms will need to be adapted for noisier image data. At this time, it is unknown how much noise can be tolerated by CTC CAD. Algorithms that reduce image noise without blurring edges (such as anisotropic diffusion [21]) may be successful at reducing the significance of noise but may be computationally intensive. Fuzzy membership rules that allow more flexible criteria for what constitutes a polyp may be needed because noise fluctuations can deviate the shape of a polyp away from the expected elliptical curvature.

Challenge 9: Low image resolution Another objective of CTC is to locate polyps in a timeefficient and cost-effective manner. For these reasons, reduced data size (thicker slices) has been used successfully for CTC [20]. Reduced data size improves efficiency and cost effectiveness because fewer images need to be inspected. One problem with reduced data size is that such methods do not produce the isotropic voxels that are desirable for three-dimensional image processing. For

273

example, it may be difficult to identify polyps that occupy only a few voxels on a single image of a reduced data size acquisition. In contrast, an acquisition with greater resolution will likely provide greater detail of polyp geometry and internal architecture that may improve automated detection. Consequently, resolution issues are likely to be significant for automated polyp detection. There are several possible solutions to the problem of reduced data size. First, the images can be interpolated to produce isotropic voxels. Second, the data can be reconstructed at different levels of resolution: a lower resolution dataset for clinical diagnosis and a higher resolution one for the polyp detector. Third, two-dimensional rather than three-dimensional image processing can be performed (or done in addition). For example, two-dimensional Hough transforms have been used to excellent effect in CTC CAD [10]. They can be applied along a narrow band of the colon wall looking for circular (polypoid) abnormalities. There may be other morphologic filters that can be applied advantageously to two-dimensional images.

Challenge 10: Integrating results of CAD into clinical practice Optimal CAD is of little value if the information provided is not used. Consequently, CAD detections must be integrated into clinical practice. Although there is some experience in the literature with integrating mammography CAD into clinical practice, very little has been reported for CTC. Possible methods for incorporating CAD into clinical practice include labeling (coloring) sites identified by CAD on axial CTC images [22] or three-dimensional perspective-rendered endoluminal surface reconstructions [23] for review by the radiologist (Figs. 1, 2). Also, automated detection could be combined with automated navigation to provide directed viewing as part of a total colonic evaluation [24 –26]. In summary, CAD for CTC has the potential to increase diagnostic sensitivity, reduce perceptual error, decrease interpretation time, increase throughput, and reduce the radiologist’s fatigue. CTC CAD in its evolution will borrow methods from other successful radiology CAD but also will require new approaches. Although there are many challenging problems to be solved, initial results have been encouraging. It promises to be an exciting area of research in the years ahead. Acknowledgments. Andrew Dwyer, M.D., is thanked for critical review of the manuscript. C. Daniel Johnson, M.D., Mayo Clinic Department of Radiology, is thanked for providing CT colonographic data. The intramural research programs of the Diagnostic Radiology Department, Warren G. Magnuson Clinical Center, supported this work. I received a U.S. Patent on the subject matter discussed in this article.

274

References 1. Fletcher JG, Johnson CD, Welch TJ, et al. Optimization of CT colonography technique: prospective trial in 180 patients. Radiology 2000;216:704 –711 2. Ferrucci JT. CT colonography for colorectal cancer screening: lessons from mammography. AJR 2000;174:1539 –1541 3. Li H, Liu KJ, Lo SC. Fractal modeling and segmentation for the enhancement of microcalcifications in digital mammograms. IEEE Trans Med Imaging 1997;16:785–798 4. 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 5. Yankelevitz DF, Reeves AP, Kostis WJ, et al. Small pulmonary nodules: volumetrically determined growth rates based on CT evaluation. Radiology 2000;217:251–256 6. Summers RM, Beaulieu CF, Pusanik LM, et al. Automated polyp detector for CT colonography: feasibility study. Radiology 2000; 216:284 –290 7. 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 8. Yoshida H, Masutani Y, MacEneaney P, et al. Computer-aided detection of polyps in CT colonography based on geometric features. Proc SPIE 2001;4321:53–57 9. Yoshida H, Masutani Y, MacEneaney PM, et al. Detection of colonic polyps in CT colonography based on geometric features. Radiology 2000;217:1555 10. 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;217:704 11. Masutani Y, Dachman AH, MacEneaney PM, et al. Computerized detection of polyps in CT colonography based on volumetric curvature information. Radiology 2000;217:1169 12. Wyatt CL, Ge Y, Vining DJ. Automatic segmentation of the colon for virtual colonoscopy. Comput Med Imaging Graph 2000;24:1–9

R. M. Summers: Challenges for CTC CAD 13. Zalis ME, Hahn PF. Digital subtraction bowel cleansing in CT colonography. AJR 2001;176:646 – 648 14. Sahiner B, Petrick N, Chan HP, et al. Recognition of lesion correspondence on two mammographic views—a new method of falsepositive reduction for computerized mass detection. Proc SPIE 2001;4322:649 – 655 15. Wang G, McFarland EG, Brown BP, et al. GI tract unraveling with curved cross sections. IEEE Trans Med Imaging 1998;17:318 –322 16. Haker S, Angenent S, Tannenbaum A, et al. Nondistorting flattening maps and the 3-D visualization of colon CT images [in process citation]. IEEE Trans Med Imaging 2000;19:665– 670 17. Paik DS, Beaulieu CF, Jeffrey RB, et al. Virtual colonoscopy visualization modes using cylindrical and planar map projections: technique and evaluation. Radiology 1998;209P:429 18. Dave SB, Wang G, Brown BP, et al. Straightening the colon with curved cross sections: an approach to CT colonography. Acad Radiol 1999;6:398 – 410 19. Zhang Z, Wang G, Brown BP, et al. Fast algorithm for soft straightening of the colon. Acad Radiol 2000;7:142–148 20. Hara AK, Johnson CD, Reed JE, et al. Reducing data size and radiation dose for CT colonography. AJR 1997;168:1181–1184 21. Gerig G, Kubler O, Kikinis R, et al. Nonlinear anisotropic filtering of MRI data. IEEE Trans Med Imaging 1992;11:221–232 22. Summers RM, Pusanik LM, Malley JD, et al. Method of labeling colonic polyps at CT colonography using computer-assisted detection. In: Lemke HV, Vannier MW, Inamura K, Farman AG, Doi K, eds. Computer assisted radiology and surgery (CARS). San Francisco: Elsevier Science, 2000:785–789 23. Summers RM. Image gallery: a tool for rapid endobronchial lesion detection and display using virtual bronchoscopy. J Digit Imaging 1998;11:53–55 24. Paik DS, Beaulieu CF, Jeffrey RB, et al. Automated flight path planning for virtual endoscopy. Med Physics 1998;25:629 – 637 25. Ge YR, Stelts DR, Vining DJ. 3D skeleton for virtual colonoscopy. Lecture Notes Comput Sci 1996;449 – 454 26. Wang J, Ge YR. An optimization problem in virtual endoscopy. Theoret Comput Sci 1998;207:203–216