(Radiographics. 2002;22:963-979.)
© RSNA, 2002
Imaging & Therapeutic Technology |
Computer-aided Diagnosis Scheme for Detection of Polyps at CT Colonography1
Hiroyuki Yoshida, PhD,
Janne Näppi, PhD,
Peter MacEneaney, MD,
David T. Rubin, MD and
Abraham H. Dachman, MD
1 From the Departments of Radiology (H.Y., J.N., P.M., A.H.D.) and Medicine (D.T.R.), University of Chicago, 5841 S Maryland Ave, MC20206, Chicago, IL 60637. Recipient of a Cum Laude award for an education exhibit at the 2000 RSNA scientific assembly. Received November 6, 2001; revision requested January 9, 2002 and final revision received April 23; accepted April 23. Supported in part by the University of Chicago Cancer Research Center and the Cancer Research Foundation of America. Address correspondence to H.Y. (e-mail: yoshida@uchicago.edu).
 |
Abstract
|
|---|
Colon cancer is one of the leading causes of cancer deaths in the United States. However, most colon cancers can be prevented if precursor colonic polyps are detected and removed. An advanced computer-aided diagnosis (CAD) scheme was developed for the automated detection of polyps at computed tomographic (CT) colonography. A region encompassing the colonic wall is extracted from an isotropic volume data set obtained by interpolating CT colonographic scans along the axial direction. Polyp candidates are detected with computation of three-dimensional (3D) geometric features that characterize polyps, followed by extraction of polyps with hysteresis thresholding and fuzzy clustering using these geometric features. The number of false-positive findings is reduced by extracting 3D texture features from polyp candidates and applying quadratic discriminant analysis to the candidates. This CAD scheme was applied in 71 patients who underwent CT colonography, 14 of whom had colonoscopically confirmed polyps (n = 21). At by-patient analysis, sensitivity was 100%, with an average false-positive rate of 2.0 per patient. At by-polyp analysis, the scheme detected 90% of the polyps at the same false-positive rate. This CAD scheme permits accurate detection of suspicious lesions and thus has the potential to reduce radiologists’ interpretation time and improve their diagnostic accuracy in the detection of polyps at CT colonography.
© RSNA, 2002
Index Terms: Colon, CT, 75.12115, 75.12117 • Colon, neoplasms, 75.311 • Computers, diagnostic aid
|
Introduction
|
|---|
Colon cancer is the second leading cause of cancer deaths in the United States, with approximately 60,000 deaths per year (1,2). Many of these deaths might be prevented with the introduction of an effective screening tool. Computed tomographic (CT) colonography is a technique used in the detection of colonic polyps, the precursors of colorectal carcinoma, with CT of the cleansed and air-distended colon (3–8). CT colonography has been advocated as a potential screening tool; however, for CT colonography to be a clinically practical means of screening for colon cancer, the technique must allow timely interpretation of a large number of images and highly accurate detection of polyps and masses.
Currently, interpretation of an entire CT colonography examination is time consuming (3,9), primarily because a single examination typically produces a total of 400–700 axial CT scans. Studies show that case interpretation time is 10–40 minutes, even with expert abdominal radiologists (3). Thus, interpretation time must be reduced substantially before CT colonography can make the transition from research to routine clinical practice, especially as a screening tool (10).
Recent advances in imaging software allow radiologists to interpret CT colonoscopy data on multiplanar reformatted images and rear endoluminal views of the colon (7,9–12). The conspicuity of polyps may depend on the display methods used; thus, the use of different views improves the detection of polyps. Several studies have evaluated polyp detection with two-dimensional (2D) axial scans alone compared with axial images in conjunction with three-dimensional (3D) endoluminal views (7,9,10). These studies show that the diagnostic quality of the combined approach is superior to that of using axial scans alone. A more recent study showed that detection with 2D multiplanar reformatted images is comparable to that with 3D endoluminal views (11).
However, regardless of whether 2D views are used alone or in conjunction with 3D views, radiologists could miss polyps and report false-positive findings due to perceptual errors caused by normal structures that mimic polyps (11,12). Therefore, the diagnostic performance of CT colonography is indeterminate, being susceptible to human error (3,13). The learning curve for the accurate interpretation of CT colonographic scans can be one of the causes for variable sensitivity among readers (14). The visibility and conspicuity of polyps, and thus the accuracy of polyp detection, may also depend on image acquisition parameters and display methods, both of which are still under investigation (15,16). These factors increase the potential for perceptual error even among experienced observers (13). A large number of CT scans for each patient and the absence of visual cues that normally exist with colonoscopy (eg, mucosal color changes) also make image interpretation tedious and susceptible to error.
To overcome these difficulties, researchers have developed preliminary computer-aided diagnosis (CAD) schemes for the automated detection of polyps at CT colonography (17–19). The "second opinion" offered by a CAD scheme has the potential to reduce radiologists’ interpretation time and to enhance diagnostic accuracy in the detection of polyps. Reduction in overall interpretation time can be achieved if radiologists focus on the small number of regions indicated as suspicious by the CAD scheme. Thus, radiologists can quickly survey a large portion of the colon that is likely to be normal. In addition, as observed in mammographic applications, CAD has the potential to reduce radiologists’ perceptual errors, thereby improving detection (20).
In this article, we describe our advanced CAD scheme for the detection of colonic polyps, which yields a high sensitivity and a low false-positive rate. We also report the evaluation of the performance of this scheme based on cases of colonoscopically confirmed polyps seen in patients who had undergone clinical CT colonography.
|
Overview of the CAD Scheme
|
|---|
Figure 1 shows a schematic for our detection scheme. First, to perform a full 3D analysis of the colon and polyps, we generate an isotropic 3D volume from axial CT scans obtained with CT colonography. Next, we "extract" the colon by determining a set of thick regions that encompasses the entire colonic wall. To detect polyp candidates, we compute two 3D geometric features for each voxel of the segmented colon. We then segment the polyp candidates with hysteresis thresholding based on these geometric features and apply fuzzy clustering to the segmented polyps to group them for identification of polyp candidates. To reduce false-positive findings, we compute additional 3D textural features that characterize the internal structure of polyps. The final detected polyps are obtained by applying discriminant analysis to the feature space generated by the geometric and textural features.
|
Extraction of Colonic Wall
|
|---|
Generation of Isotropic Volume
For full 3D analysis of the colon and polyps, an isotropic 3D volume is generated from the axial CT scans in a CT colonography data set. The axial CT scans are interpolated linearly along the axial direction so that the difference between the spatial and axial resolution in a single data set is eliminated. A typical CT colonography data set obtained with the patient in either the supine or prone position consists of 200–350 axial CT scans with a 512 x 512 matrix. The spatial resolution (in-plane pixel dimension) varies among patients and ranges from 0.5 to 0.75 mm per voxel. Therefore, after interpolation, the height, width, and depth of the isotropic volume are typically 512, 512, and 500–850, respectively. In this article, we refer to these 3D isotropic volumetric data simply as isotropic volume (Fig 2).
View larger version (97K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2a. Generation of an isotropic volume for CT colonography. Axial CT scans are obtained from a CT colonography examination (a), and an isotropic volume is generated with linear interpolation between the CT scans along the axial direction (b).
|
|
View larger version (117K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2b. Generation of an isotropic volume for CT colonography. Axial CT scans are obtained from a CT colonography examination (a), and an isotropic volume is generated with linear interpolation between the CT scans along the axial direction (b).
|
|
Knowledge-guided Colon Extraction
Computer analysis of the entire isotropic volume would be inefficient because the computer would spend a significant amount of time analyzing extracolonic structures where polyps are unlikely to occur. In addition, extracolonic structures may generate a large number of false-positive results. Therefore, we developed a fully automated two-step method for extracting the entire colonic wall from the isotropic volume.
In the first step, called anatomy-based extraction, we extract a set of thick regions, which encompasses the entire colonic wall, based on a priori knowledge of abdominal anatomy. The resulting binary volume, called a colon mask, indicates the voxels that belong to the extracted set of regions. In anatomy-based extraction, we first segment and remove the following extracolonic anatomic structures from the isotropic volume: (a) the area surrounding the body region ("outer air"), (b) the osseous structures (the spine, pelvis, and parts of the ribs), and (c) the lung bases. For this purpose, we apply a thresholding process to the isotropic volume based on CT values, which extracts regions with a range of CT values. We then perform connected component analysis, which yields the contiguous regions corresponding to the extracolonic structures (Fig 3). The threshold values are determined adaptively from the histogram of the isotropic volume by identifying the characteristic peaks corresponding to air, fat, and muscle (21). The extracted components are subjected to 3D morphologic dilation (22), so that each extracolonic structure can be removed completely (Fig 4). After these components have been removed, the colonic wall is extracted with thresholding of the remaining volume based on the CT and gradient magnitude values. The segmented regions are subjected to connected component analysis, and the resulting components whose volume exceeds 1% of the volume of the largest component are identified collectively as the colon mask (Fig 5).
View larger version (80K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4a. Extracolonic structures that have been extracted from the thresholded isotropic volume shown in Figure 3: the air surrounding the body region ("outer air") (a), the osseous structures (spine, pelvis, and parts of the ribs) (b), and the lung bases (c). (Fig 4a-4c reprinted, with permission, from reference 19.)
|
|
View larger version (83K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4b. Extracolonic structures that have been extracted from the thresholded isotropic volume shown in Figure 3: the air surrounding the body region ("outer air") (a), the osseous structures (spine, pelvis, and parts of the ribs) (b), and the lung bases (c). (Fig 4a-4c reprinted, with permission, from reference 19.)
|
|
View larger version (22K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4c. Extracolonic structures that have been extracted from the thresholded isotropic volume shown in Figure 3: the air surrounding the body region ("outer air") (a), the osseous structures (spine, pelvis, and parts of the ribs) (b), and the lung bases (c). (Fig 4a-4c reprinted, with permission, from reference 19.)
|
|
View larger version (85K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 5. Colon mask extracted at the preprocessing step of the CAD scheme. The extracted colon mask is a thick region that encompasses the entire colonic wall, not just the surface of the wall.
|
|
An advantage of this method over surface generation methods (23–25) is that we can extract entire polyps in the colonic wall, including their internal structure. A further advantage is that the method can reliably extract the entire colon, even when it is collapsed in several places (Fig 6). In a previous study, visual inspection by radiologists confirmed that our method was capable of extracting all parts of a collapsed colon in a preliminary database (21).
View larger version (63K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 6. Colon mask extracted from a colon that has collapsed in several places. The CAD scheme extracts the entire colon despite its collapse at the end of the sigmoid colon (bottom arrow) and in the middle of transverse colon (top arrow).
|
|
The extracted colon mask may contain small bowel, stomach, or both that adhere to the wall of the colon because of their spatial contiguity with the colonic wall. It is desirable to remove them because residual materials in the small bowel and stomach can cause false-positive results with the CAD scheme. To this end, we implemented a second step called colon-based analysis, which is based on automated selection of "seed points" in the colon mask obtained with anatomy-based extraction. In colon-based analysis, we first locate the largest connected component in the center region of the bottom one-tenth of the entire colon mask. This component is identified as the end of the rectum. We designate the voxel with the low-est CT value as a seed point. The volume-growing process is initiated from this seed point. Once the colonic lumen is segmented with this process, new surfaces are added to the volume-grown region to encompass the colonic wall. 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 volume-grown region and the colon mask with anatomy-based extraction removes the small bowel and stomach and determines the final colon mask, which encompasses only the colon. The process for the removal of small bowel with colon-based analysis is illustrated with 2D multiplanar reformatted images in Figure 7 and with 3D volume-rendered views in Figure 8.
View larger version (115K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7a. Removal of small bowel with colon-based analysis. Coronal multiplanar reformatted images demonstrate a portion of an isotropic volume (a), a colon mask (green) extracted with the anatomy-based approach (b), the small bowel (red) removed with colon-based analysis (c), and the final colon mask (green) extracted with colon-based analysis (d).
|
|
View larger version (143K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7b. Removal of small bowel with colon-based analysis. Coronal multiplanar reformatted images demonstrate a portion of an isotropic volume (a), a colon mask (green) extracted with the anatomy-based approach (b), the small bowel (red) removed with colon-based analysis (c), and the final colon mask (green) extracted with colon-based analysis (d).
|
|
View larger version (141K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7c. Removal of small bowel with colon-based analysis. Coronal multiplanar reformatted images demonstrate a portion of an isotropic volume (a), a colon mask (green) extracted with the anatomy-based approach (b), the small bowel (red) removed with colon-based analysis (c), and the final colon mask (green) extracted with colon-based analysis (d).
|
|
View larger version (142K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7d. Removal of small bowel with colon-based analysis. Coronal multiplanar reformatted images demonstrate a portion of an isotropic volume (a), a colon mask (green) extracted with the anatomy-based approach (b), the small bowel (red) removed with colon-based analysis (c), and the final colon mask (green) extracted with colon-based analysis (d).
|
|
|
Detection of Polyp Candidates
|
|---|
After the colon mask is extracted, polyp candidates in the colonic wall are detected by extracting geometric features that effectively characterize polyps at each point in the wall. Polyps adhering to the colonic wall tend to appear as relatively small, bulbous, caplike structures, whereas folds appear as elongated, ridgelike structures (Fig 9). The colonic wall itself appears as a large, nearly flat cuplike structure. To characterize these shape and scale differences among polyps, folds, and colonic wall, we use two 3D geometric features called the volumetric shape index and volumetric curvedness (26–28).
View larger version (39K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 9. Geometric model for structures in the colonic lumen. Polyps tend to appear as bulbous, caplike structures that adhere to the colonic wall; folds appear as elongated, ridgelike structures; and the colonic wall appears as a large, nearly flat cuplike structure.
|
|
The volumetric shape index characterizes the topologic shape of the volume in the vicinity of a voxel. The volumetric shape index can be defined at every point in a volume (29) and captures the intuitive notion of the local shape of the surface at each point. Every distinct shape class has a unique shape index value (26,29); these values range from 0 to 1. For example, five well-known shape classes have the following shape index values: "cup" (0.0), "rut" (0.25), "saddle" (0.5), "ridge" (0.75), and "cap" (1.0) (Fig 10). One of the most important advantages of using the shape index in shape analysis is that the transition from one shape class to another is continuous; thus, the shape index can be used to describe subtle shape variations effectively. For example, a shape index value of 0.875 represents the "dome" class, which is a transitional shape class between ridge (0.75) and cap (1.0).
View larger version (53K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 10. Relationship between volumetric shape index values and shape classes. Voxels that belong to the cup, rut, saddle, ridge, and cap classes have shape index values of 0, 0.25, 0.5, 0.75, and 1, respectively. An intermediate shape index value represents a transitional shape class. For example, a shape index value of 0.875 represents the "dome" class, which is transitional between the ridge class (0.75) and the cap class (1.0). (Reprinted, with permission, from reference 18.)
|
|
Volumetric curvedness represents how gently curved a surface is (Fig 11) and is defined in terms of the magnitude of the effective curvature on the surface. The dimension of curvedness is the reciprocal of length and ranges from negative infinity to positive infinity. It measures "how much" shape the neighborhood of a voxel includes, whereas the shape index measures "which" shape the local neighborhood of the voxel has. Curvedness also provides scale information about an object: A large negative value implies a very gentle change, whereas a large positive value implies a very sharp knifelike edge (ie, abrupt change). Generally, points on a large spheric object have small curvedness values. The colonic wall, polyps, and folds have small, medium, and large curvedness values, respectively. Therefore, a combination approach that makes use of both the shape index and curvedness can help differentiate effectively among these three "objects."
Figure 12 demonstrates the effectiveness of the shape index in differentiating polyps from normal structures. Each of the four pairs of images (Fig 12a–Fig 12d) consists of a 2D multiplanar reformatted image of a region that includes a polyp (left) along with the corresponding 3D volume-rendered endoscopic view (right). As shown on the color bar (Fig 12e), voxels that have shape index values corresponding to the cap class are colored green, those corresponding to the ridge class are colored pink, and those corresponding to other classes are colored brown. As expected, a substantial portion of the polyp is colored green, whereas folds and colonic walls are colored pink and brown, respectively. With this color scheme, the polyps, folds, and colonic wall are clearly delineated, and the polyp is easily distinguishable from other structures.
View larger version (62K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 12a. Use of the shape index feature in the characterization of polyps. (a-d) CT scans (bottom) and their corresponding 3D volume-rendered endoscopic views (top) demonstrate polyps (arrows). (Fig 12a and Fig 12c reprinted, with permission, from references 19 and 18, respectively.) (e) Color bar demonstrates that voxels with shape index values corresponding to the cap class, saddle-ridge class, and other classes are colored green, pink, and brown, respectively. With this color scheme, the polyps, folds, and colonic wall are clearly delineated, and the polyp is easily distinguishable from other structures.
|
|
View larger version (60K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 12b. Use of the shape index feature in the characterization of polyps. (a-d) CT scans (bottom) and their corresponding 3D volume-rendered endoscopic views (top) demonstrate polyps (arrows). (Fig 12a and Fig 12c reprinted, with permission, from references 19 and 18, respectively.) (e) Color bar demonstrates that voxels with shape index values corresponding to the cap class, saddle-ridge class, and other classes are colored green, pink, and brown, respectively. With this color scheme, the polyps, folds, and colonic wall are clearly delineated, and the polyp is easily distinguishable from other structures.
|
|
View larger version (56K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 12c. Use of the shape index feature in the characterization of polyps. (a-d) CT scans (bottom) and their corresponding 3D volume-rendered endoscopic views (top) demonstrate polyps (arrows). (Fig 12a and Fig 12c reprinted, with permission, from references 19 and 18, respectively.) (e) Color bar demonstrates that voxels with shape index values corresponding to the cap class, saddle-ridge class, and other classes are colored green, pink, and brown, respectively. With this color scheme, the polyps, folds, and colonic wall are clearly delineated, and the polyp is easily distinguishable from other structures.
|
|
View larger version (66K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 12d. Use of the shape index feature in the characterization of polyps. (a-d) CT scans (bottom) and their corresponding 3D volume-rendered endoscopic views (top) demonstrate polyps (arrows). (Fig 12a and Fig 12c reprinted, with permission, from references 19 and 18, respectively.) (e) Color bar demonstrates that voxels with shape index values corresponding to the cap class, saddle-ridge class, and other classes are colored green, pink, and brown, respectively. With this color scheme, the polyps, folds, and colonic wall are clearly delineated, and the polyp is easily distinguishable from other structures.
|
|
View larger version (9K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 12e. Use of the shape index feature in the characterization of polyps. (a-d) CT scans (bottom) and their corresponding 3D volume-rendered endoscopic views (top) demonstrate polyps (arrows). (Fig 12a and Fig 12c reprinted, with permission, from references 19 and 18, respectively.) (e) Color bar demonstrates that voxels with shape index values corresponding to the cap class, saddle-ridge class, and other classes are colored green, pink, and brown, respectively. With this color scheme, the polyps, folds, and colonic wall are clearly delineated, and the polyp is easily distinguishable from other structures.
|
|
To identify polyp candidates, we use a range of shape index and curvedness values that characterize polyps. First, voxels that have shape index values between 0.9 and 1.0 and curvedness values between the effective sizes of 5.0 and 12.5 mm are extracted for selection of seed regions that are in the cap class and in the targeted polyp size ranges. These predefined threshold values for shape index and curvedness help identify small portions of polyps that have a nearly perfect cap shape.
Hysteresis thresholding based on the shape index is applied, starting with seed points, to the colonic wall for extraction of polyp candidates (Fig 13a, Fig 13b) (22). This process extracts a set of voxels that are spatially connected to the seed points and have shape index values within a predefined range. Hysteresis thresholding is used to extract a large connected component that corresponds to the major portion of a polyp. This is necessary because the peripheral region of a polyp does not always have a perfect cap shape, but may demonstrate a domelike shape instead. The peripheral region may also have curvedness values that differ from those of the central region. Therefore, we use a range of threshold values between 0.8 and 1.0 for the shape index and effective sizes of 4–20 mm for curvedness to extract a significant portion of a polyp. Figure 13c shows an extracted region obtained with hysteresis thresholding that corresponds to the polyp shown in Figure 13a.
View larger version (117K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 13a. Extraction of polyp candidates with hysteresis thresholding. Sagittal CT scans show a polyp (arrow) within the region of interest (a), the segmented colonic wall (blue) and the seed region for the hysteresis thresholding (red) (b), and an extracted region obtained with hysteresis thresholding that corresponds to the polyp (green) (c).
|
|
View larger version (58K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 13b. Extraction of polyp candidates with hysteresis thresholding. Sagittal CT scans show a polyp (arrow) within the region of interest (a), the segmented colonic wall (blue) and the seed region for the hysteresis thresholding (red) (b), and an extracted region obtained with hysteresis thresholding that corresponds to the polyp (green) (c).
|
|
View larger version (134K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 13c. Extraction of polyp candidates with hysteresis thresholding. Sagittal CT scans show a polyp (arrow) within the region of interest (a), the segmented colonic wall (blue) and the seed region for the hysteresis thresholding (red) (b), and an extracted region obtained with hysteresis thresholding that corresponds to the polyp (green) (c).
|
|
A polyp candidate may generate multiple detections at different locations within the polyp due to multiple small, bumpy structures on its surface. Therefore, we use a fuzzy c-means clustering algorithm that merges multiple detections of the same polyp (30). This algorithm groups data points (ie, voxels in polyp candidates) with similar feature values into a single cluster. Similarity is defined in terms of the distance between the feature values of each data point. Therefore, a major portion of a polyp is extracted as a large cluster because the voxels within a single polyp are expected to have similar features and to be located close together. Moreover, multiple detections at different locations of a single polyp can be merged into a single, large, possibly unconnected cluster with fuzzy clustering.
Fuzzy clustering maintains polyp candidates that are due to noise as a small, isolated cluster because they tend to contain voxels that have feature values distinctly different from those of the surrounding voxels. Therefore, thresholding with a minimum volume of 35 mm3 (equivalent to the volume of a 4-mm polyp) is applied to individual clusters to remove these small noise-related clusters while retaining clinically significant polyps that are 5 mm or larger.
|
Reduction of False-Positive Findings
|
|---|
Directional Gradient Concentration
Polyp candidates obtained in the detection process consist not only of polyps, but also of normal structures similar to polyps (ie, false-positive findings). Because a low false-positive rate is critical if CAD is to serve as an effective screening tool, we developed a method of reducing the number of false-positive findings among the polyp candidates obtained with fuzzy clustering. This method makes use of a 3D textural feature called the directional gradient concentration (DGC), which is a variant of the gradient concentration feature (31). We use the DGC feature in combination with the shape index and variance of CT values for efficient differentiation of true-positive from false-positive findings. A gradient vector at a voxel represents the rate of change in the CT value at that point. In particular, the direction of the gradient vector indicates the direction of the change. In the vicinity of a given point, the gradient concentration characterizes the directions of the gradient vectors relative to that point. If most gradient vectors are directed toward the point, the gradient concentration feature has a high value; otherwise, it has a low value.
Because of partial volume effect, the CT value of soft tissue within a small polyp tends to increase as one moves from the colonic air toward the center of the polyp. Therefore, gradient vectors that converge at the center of the polyp tend to appear at voxels near the boundary between the polyp and the colonic air. Because small polyps are generally hemispheric objects on the colonic wall, the gradient vectors pointing to the center of the polyp tend to form a hemispheric concentration pattern. However, the degree of gradient concentration at a given point is generally highest when the surrounding gradient vectors form a spheric concentration pattern. Therefore, we defined the DGC feature from the gradient concentration by subtracting the gradient concentration computed from opposite directions at a voxel. As a result, the DGC has the highest value for hemispheric gradient concentration patterns, whereas it has a low value for spheric concentration patterns.
The DGC features can be used effectively for differentiation of polyps from folds and stool. Folds are elongated, ridgelike objects and thus have no center toward which the CT value increases. Therefore, the gradient vectors in a fold do not concentrate toward any particular point. Stool tends to exhibit an inhomogeneous internal texture pattern and thus causes the gradient vectors to be directed randomly rather than toward any particular point. Figure 14 demonstrates how use of the gradient concentration can highlight a polyp and effectively help differentiate it from a fold.
View larger version (58K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 14a. Use of the DGC feature in the characterization of polyps. (a) Three-dimensional object demonstrates a simulated fold and polyp. (b) A cut-plane volume-rendered view of the simulated fold and polyp is color-coded according to the color scheme for DGC values. The fold and polyp are shown with relatively low (pink-red) and high (green) DGC values, respectively. (c) Axial CT scan demonstrates a 9-mm polyp (arrow). (d, e) Volume-rendered (d) and color-coded (e) endoscopic views more clearly depict the polyp shown in c. (f) Color bar demonstrates the color scheme for DGC values.
|
|
View larger version (87K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 14b. Use of the DGC feature in the characterization of polyps. (a) Three-dimensional object demonstrates a simulated fold and polyp. (b) A cut-plane volume-rendered view of the simulated fold and polyp is color-coded according to the color scheme for DGC values. The fold and polyp are shown with relatively low (pink-red) and high (green) DGC values, respectively. (c) Axial CT scan demonstrates a 9-mm polyp (arrow). (d, e) Volume-rendered (d) and color-coded (e) endoscopic views more clearly depict the polyp shown in c. (f) Color bar demonstrates the color scheme for DGC values.
|
|
View larger version (93K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 14c. Use of the DGC feature in the characterization of polyps. (a) Three-dimensional object demonstrates a simulated fold and polyp. (b) A cut-plane volume-rendered view of the simulated fold and polyp is color-coded according to the color scheme for DGC values. The fold and polyp are shown with relatively low (pink-red) and high (green) DGC values, respectively. (c) Axial CT scan demonstrates a 9-mm polyp (arrow). (d, e) Volume-rendered (d) and color-coded (e) endoscopic views more clearly depict the polyp shown in c. (f) Color bar demonstrates the color scheme for DGC values.
|
|
View larger version (138K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 14d. Use of the DGC feature in the characterization of polyps. (a) Three-dimensional object demonstrates a simulated fold and polyp. (b) A cut-plane volume-rendered view of the simulated fold and polyp is color-coded according to the color scheme for DGC values. The fold and polyp are shown with relatively low (pink-red) and high (green) DGC values, respectively. (c) Axial CT scan demonstrates a 9-mm polyp (arrow). (d, e) Volume-rendered (d) and color-coded (e) endoscopic views more clearly depict the polyp shown in c. (f) Color bar demonstrates the color scheme for DGC values.
|
|
View larger version (100K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 14e. Use of the DGC feature in the characterization of polyps. (a) Three-dimensional object demonstrates a simulated fold and polyp. (b) A cut-plane volume-rendered view of the simulated fold and polyp is color-coded according to the color scheme for DGC values. The fold and polyp are shown with relatively low (pink-red) and high (green) DGC values, respectively. (c) Axial CT scan demonstrates a 9-mm polyp (arrow). (d, e) Volume-rendered (d) and color-coded (e) endoscopic views more clearly depict the polyp shown in c. (f) Color bar demonstrates the color scheme for DGC values.
|
|
View larger version (10K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 14f. Use of the DGC feature in the characterization of polyps. (a) Three-dimensional object demonstrates a simulated fold and polyp. (b) A cut-plane volume-rendered view of the simulated fold and polyp is color-coded according to the color scheme for DGC values. The fold and polyp are shown with relatively low (pink-red) and high (green) DGC values, respectively. (c) Axial CT scan demonstrates a 9-mm polyp (arrow). (d, e) Volume-rendered (d) and color-coded (e) endoscopic views more clearly depict the polyp shown in c. (f) Color bar demonstrates the color scheme for DGC values.
|
|
Quadratic Discriminant Analysis
We use statistical classifiers to combine the 3D geometric and textural features to obtain the final detected polyps. For effective classification, we use a classifier based on quadratic discriminant analysis (QDA) (32). This classifier combines the features into a single index as follows: On the basis of training samples that are labeled as either true-positive or false-positive findings, QDA generates a decision boundary that optimally partitions the feature space (consisting of feature values for the polyp candidates) into a "polyp" class and a "false-positive" class. QDA uses a hyperquadratic surface for the decision boundary. For a polyp candidate, the distance from the decision boundary, called polyp likelihood, provides the ranked ordering of the likelihood that the candidate is a polyp or a false-positive finding. A polyp candidate is classified as a polyp if its polyp likelihood is higher than a predefined decision threshold value. Otherwise, it is classified as a false-positive finding. The polyp candidates that are classified as polyps constitute the final detected polyps as reported with the CAD scheme.
In a previous study, we evaluated the performance of various features in discriminating between true-positive and false-positive findings when combined by the QDA-based classifier (31). We found that combining the shape index, gradient concentration, DGC, CT values, and gradient and variance of CT values was particularly effective in this setting. Figure 15 shows the distribution of the mean shape index values and mean DGC values for the polyp candidates obtained with fuzzy clustering from supine and prone data sets in all patients. The solid curve shows a decision boundary generated by the QDA-based classifier that effectively separates true-positive from false-positive findings. The final detected polyps are then displayed on a graphical user interface (Fig 16).
View larger version (55K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 15. Distribution of feature values for polyp candidates obtained with fuzzy clustering. Graph illustrates the distribution of the mean shape index values and mean DGC values for the polyp candidates obtained from supine and prone data sets in all patients. Solid curve represents a decision boundary generated by a QDA-based classifier. Black squares = true-positive findings, gray dots = false-positive findings.
|
|
View larger version (94K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 16. Graphical user interface for our CAD scheme for the detection of polyps. A list of the locations of detected polyps is displayed in the small window in the lower left corner. By clicking on an entry in the list, one can quickly jump to an image of the suspected polyp for visual confirmation. The polyp candidates are shown on three multiplanar reformatted images of the CT colonography data. A color-coded endoscopic view of the suspected polyp (green) is displayed for 3D problem solving as needed.
|
|
|
Detection Performance in Clinical Cases
|
|---|
Patients and Procedures
CT colonography studies of 71 patients (31 men and 40 women) were selected from 100 CT colonography examinations performed at our institution between 1997 and 2001. Patients ranged in age from 32 to 84 years (mean, 59 years). After the standard precolonoscopy cleansing, all patients underwent CT colonography, followed by colonoscopy on the same day. Two experienced radiologists independently examined the resulting images visually. Cases in which there were feces, fluid, a poorly distended colon, or motion artifacts were included if they were of diagnostic quality despite these artifacts. Cases with colitis were excluded. Helical CT was performed with GE 9800 CTi and LightSpeed QX/i scanners (GE Medical Systems, Milwaukee, Wis) with the patient in both the supine and prone positions after rectal insufflation with room air. Scanning parameters were as follows: 2.5–5-mm collimation, 1.5–2.5-mm reconstruction intervals, 512 x 512 matrix for the resulting axial images, 0.5–0.7-mm/pixel spatial resolution, and a reduced current of 60 or 100 mA with 120 kVp to minimize radiation exposure.
This cohort of 71 patients included 14 patients with colonic polyps and 57 patients without polyps. The former group had a total of 21 colonoscopically confirmed polyps that were 5 mm or larger. These polyps were located in the sigmoid colon (n = 8), ascending colon (n = 5), hepatic flexure (n = 3), splenic flexure (n = 2), descending colon (n = 1), transverse colon (n = 1), and cecum (n = 1). The exact locations of these polyps were determined with visual confirmation at CT colonography by the two radiologists, who had access to the colonoscopy and pathology reports and were able to consult with an endoscopist.
Of the 21 polyps, 15 were at least 5 mm but less than 10 mm in diameter. The remaining six polyps were 10 mm or more in diameter; one polyp measured 25 mm, and the other five polyps were less than 20 mm. The size of the polyps was determined primarily on the basis of the colonoscopy and pathology reports. We defined a clinically significant polyp as being at least 5 mm in diameter, which is the lower limit for clinically significant polyps (3). We chose this lower limit as a conservative criterion because the measured size of polyps varies depending on their conspicuity.
Detection Performance of the CAD Scheme
The locations of the polyps as detected with the CAD scheme were compared with their true locations. The detected polyps that were within 10 mm of their actual location were identified as true-positive findings, whereas all others were identified as false-positive findings. The CAD scheme processed the supine and prone volumetric data sets independently to yield polyp candidates. Two methods were used to evaluate detection performance: (a) by-patient analysis, in which a case was regarded as abnormal if at least one true polyp was detected in either the supine or prone data set, and (b) by-polyp analysis, in which a polyp was regarded as detected if it was detected in either the supine or prone data set of a single patient. In both methods, the average number of false-positive findings per patient was calculated as an index of the false-positive rate.
To evaluate the unbiased performance of the QDA-based classifier, we used a round-robin (or "leave-one-out") evaluation method (33–35). In this method, each of the polyp candidates was removed in turn from the set of all polyp candidates, and the classifier was trained using the remaining candidates to generate a decision boundary. The polyp likelihood for the removed candidate was then calculated based on this decision boundary. This process was repeated until each candidate had served as the "left-out" candidate and its polyp likelihood calculated. We generated a free-response receiver operating characteristic (ROC) curve that indicates the overall performance with use of polyp likelihood as the sweeping variable (36). A free-response ROC curve plots the sensitivity of the CAD scheme as a function of the average number of false-positive findings per patient (Fig 17). Our CAD scheme yielded 100% sensitivity with 2.0 false-positive findings per patient at by-patient analysis and 90% sensitivity (19 of 21 polyps) at the same false-positive rate at by-polyp analysis (Table).
View larger version (24K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 17. Graph illustrates free-response ROC curves representing the detection performance of the CAD scheme. The solid line and dotted line represent detection performance based on by-patient and by-polyp analysis, respectively. For both curves, the average number of false-positive findings per patient was used as an index of the false-positive rate.
|
|
Analysis of False-Negative Findings
Two polyps were missed at by-polyp analysis. One was an 8-mm polyp that was located at a narrow valley where two folds merge, which distorted the shape of the polyp (Fig 18). Moreover, motion artifact made the polyp appear blurred. Consequently, the polyp did not contain adequate voxels with high shape index values and was not detected. The other was a 6-mm polyp that appeared to be flat and much smaller than expected, probably due to partial volume effect. The false-negative findings occurred in cases in which other polyps were detected with the CAD scheme. Therefore, the sensitivity of by-patient analysis was 100%.
View larger version (117K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 18a. False-negative finding. (a) Axial CT scan shows the region of interest (yellow box). (b) Magnified view of the region of interest shows a polyp (arrow) that was missed with the CAD scheme.
|
|
View larger version (67K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 18b. False-negative finding. (a) Axial CT scan shows the region of interest (yellow box). (b) Magnified view of the region of interest shows a polyp (arrow) that was missed with the CAD scheme.
|
|
Analysis of False-Positive Findings
The types of false-positive findings included in our results were similar to those encountered by radiologists (14,37). They all had high shape index values and low curvedness values and thus exhibited polyplike shapes. However, most of these false-positive findings could easily be distinguished from true polyps by experienced radiologists with use of both supine and prone CT colonographic views.
Approximately 40% of the false-positive findings detected with the CAD scheme were caused by folds (Fig 19), 20% by retained stool (Fig 20), 15% by residual materials inside the small bowel and stomach (Fig 21), 10% by ileocecal valves (Fig 22), and 5% by flexural pseudotumors. Other causes of false-positive findings included a rectal tube inserted through the anus for air insufflation during CT colonography, elevation of the anorectal junction by the rectal tube, and motion artifacts. Diverticula caused no false-positive findings because our computerized detection method is designed to detect only caplike structures, not cuplike structures.
View larger version (80K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 19. False-positive finding due to a prominent fold. On an axial CT scan, the tip of the fold (arrow) appears to be a caplike structure, which led to its being incorrectly identified as a polyp. Other types of folds that can create false-positive results include sharp folds at the sigmoid colon, prominent folds on the colonic wall, two converging folds, ends of folds in a tortuous colon, and folds in a poorly distended colon.
|
|
View larger version (97K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 20a. False-positive finding due to retained stool in the colon. Stool is often a major source of error for radiologists as well as for the CAD scheme when it has a caplike appearance and thus mimics polyps. (a) Axial CT scan demonstrates what appears to be a polyp (arrow). (b) Three-dimensional volume-rendered endoscopic image is color-coded according to the shape index color scheme shown in Figure 12 | |
Top
Abstract
Introduction
