Radiology
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

DOI: 10.1148/radiol.2331031326
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Summers, R. M.
Right arrow Articles by Johnson, C. D.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Summers, R. M.
Right arrow Articles by Johnson, C. D.
(Radiology 2004;233:266-272.)
© RSNA, 2004

Technical Developments

CT Colonography with Computer-aided Detection: Automated Recognition of Ileocecal Valve to Reduce Number of False-Positive Detections1

Ronald M. Summers, MD, PhD, Jianhua Yao, PhD and C. Daniel Johnson, MD

1 From the Department of Diagnostic Radiology, 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., J.Y.); and Department of Radiology, Mayo Clinic, Rochester, Minn (C.D.J.). Received August 22, 2003; revision requested October 31; revision received December 16; accepted January 30, 2004. Supported by the intramural research programs of the Diagnostic Radiology Department, Warren G. Magnuson Clinical Center. Supported in part by NIH grant RO1CA75333. Address correspondence to R.M.S. (e-mail: rms@nih.gov).


   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 Materials and Methods
Results
 Discussion
 REFERENCES
 
The ileocecal valve (ICV) is a common cause of false-positive detections of polyps at computed tomographic (CT) colonography with computer-aided detection (CAD). The authors developed a CAD algorithm for differentiating the ICV from a true polyp and evaluated this algorithm by using two colonoscopy-confirmed CT colonography data sets. Data sets 1 and 2 consisted of the data obtained at CT colonographic examinations performed in 20 and 40 patients, respectively. Forty of these patients had at least one polyp 1 cm or larger. For data set 1, the proposed ICV recognition algorithm eliminated three of nine (33%; 95% confidence interval [CI]: 8%, 70%) false-positive CAD detections that were attributable to the ICV and none of the true-positive polyp detections. For data set 2, with use of identical parameters, the algorithm eliminated 11 of 18 (61%; 95% CI: 36%, 83%) false-positive detections that were attributable to the ICV and none of the true-positive detections. The thresholds used to recognize the ICV were a mean internal CT attenuation of less than –124 HU and a volume of greater than 1.5 cm3. The proposed algorithm successfully recognized the ICV and eliminated it in some cases. This result is clinically important because, by reducing the frequency of a common cause of false-positive detections, this algorithm may improve the efficiency of physicians who use CAD.

© RSNA, 2004

Index terms: Colon, CT, 75.12115, 75.12117 • Colon neoplasms, 75.311 • Colon neoplasms, CT, 75.12115, 75.12117 • Computers, diagnostic aid


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
Materials and Methods
Results
 Discussion
 REFERENCES
 
Computer-aided detection (CAD) has been proposed as a means of improving the consistency of computed tomographic (CT) colonography interpretations and increasing sensitivity. Research at a number of academic centers supports the notion that CT colonography with CAD is feasible (1).

Among the several areas of research in CT colonography with CAD, one of the most active is that aimed at reducing the number of false-positive detections (2,3). The consequence of false-positive detections is the inappropriate referral of the patient for a needless, more invasive conventional colonoscopic examination.

Common causes of false-positive detections at CT colonography with CAD include normal haustral folds, the ileocecal valve (ICV), the enema tip, and stool. In preliminary CAD experiments, we have found false-positive detections that are due to the detection of the ICV to be particularly troublesome because these cases occur frequently and are time-consuming to interpret and exclude. The purpose of this study was to evaluate a CAD algorithm for recognition of the ICV as distinct from a true polyp.


   Materials and Methods
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
Results
 Discussion
 REFERENCES
 
Patient Population
The study cohort consisted of 60 patients who have been described previously (4,5) and whose results were grouped into two data sets. Data set 1 consisted of the data obtained in 20 patients (10 men, 10 women; mean age ± standard deviation, 65 years ± 11; age range, 40–83 years) with known polyps that were discovered at barium enema examination or sigmoidoscopy (4). Data set 2 consisted of the data obtained in 40 asymptomatic patients (19 men, 21 women; mean age, 63 ± 7 years; age range, 51–75 years) who were at high risk for colorectal cancer: 20 patients who had at least one polyp 1 cm or larger at conventional colonoscopy performed on the same day as CT colonography and 20 consecutive patients with normal results of colonoscopy performed on the same day as CT colonography (5). Patients were considered to be at high risk on the basis of a family or personal history of colorectal cancer; therefore, these patients were part of a surveillance population. All patients underwent complete colonoscopy extending to the cecum. This study was approved by the institutional review board of Mayo Clinic, and informed consent was obtained from all patients.

CT Scanning
Patients underwent standard oral colonoscopic preparation—that is, by drinking 1 gallon of polyethylene glycol electrolyte solution (Colyte; Reed and Carnrick, Jersey City, NJ) and taking two 5-mg bisacodyl tablets (Dulcolax; CIBA-Consumer, Edison, NJ)—and CO2 insufflation of the colon according to tolerance. In all patients in the data set 2 group and in one patient in the data set 1 group, 1 mg of glucagon (Eli Lilly, Indianapolis, Ind) was administered subcutaneously 10 minutes before CT scanning.

The CT colonographic examinations were performed as follows: In all 20 data set 1 patients and in one data set 2 patient, by using a CT HiSpeed Advantage single–detector row scanner; in 36 data set 2 patients, by using a LightSpeed QX/i four–detector row scanner; and in three data set 2 patients, by using a LightSpeed Plus four–detector row scanner. (All scanners were manufactured by GE Medical Systems, Milwaukee, Wis.)

CT scanning parameters were 120 kVp, 70 mAs (mean), 5-mm collimation, and a 3-mm reconstruction interval with a 2-mm overlap (6). Patients were scanned while they were prone and while they were supine. Supine scanning was performed first and followed by prone scanning. The size of the typical CT colonography data set (supine or prone) was approximately 80 MB (160 images).

Colonoscopy
Conventional colonoscopy was performed after CT colonography on the same day by a team of experienced staff colonoscopists. The colonoscopists were not aware of the CT colonographic results. Polyp sizes were determined by the colonoscopist at the time of the examination by using a probe or forceps. For purposes of brevity, polyps 1.0 cm or larger and polyps smaller than 1.0 cm are hereafter referred to as large and small polyps, respectively. Because the CAD system was trained to find mainly large polyps, small polyps are discussed no further.

Identifying the ICV and Data Recording
The location of each ICV was recorded in a data file. At both the supine and the prone CT colonographic examinations, a single radiologist (R.M.S.) identified the ICV on transverse images by clicking on a voxel inside (approximately in the center of) the ICV; only that the voxel was within the ICV—not the precise voxel—was important. In the majority of cases, the ICV could be readily identified by locating the terminal ileum and tracing it to its communication with the cecum.

The data file of voxels, which consisted of 1 voxel per ICV per CT colonographic examination, was downloaded to a prototype polyp segmenter (7). This segmenter identified the edges of polyps, masses, and ICVs by using an iterative process involving a knowledge-based deformable contour and fuzzy clustering. The segmenter reported the total volume (in cubic centimeters) and the mean interior attenuation (in Hounsfield units) of an object. The mean interior attenuation is the average attenuation of voxels within—but not at the boundary of—the segmented region.

The segmenter produced a list of ICVs that it could analyze and a list of ICVs that were unlikely to represent polyps. The findings that the segmenter could not segment were eliminated. Several criteria had to be met before a detected ICV could be deemed segmentable: First, the boundary between the ICV region and the lumen had to be convex—that is, at least one-fourth of the boundary pixels had to have a curvature smaller than –0.5 cm–1. Second, the ICV region had to be large enough: more than 8 pixels. Third, the ICV region could not be flat (width/height ratio of <3.5 for nonflat region). These criteria were derived from our polyp segmentation algorithm. ICVs that cannot be segmented are not likely to be detected as false-positive polyps.

We recorded results in two ways: First, we analyzed all known ICVs to determine how many of them could be eliminated and at which processing stage. Second, we ran the CAD algorithm and determined how many of the false-positive CAD marks were on an ICV and how many of these marks could be eliminated. The CAD algorithm is discussed in the following text.

The rationale for using the described ICV recognition algorithm is the clinical imaging observation that the ICV is large and has low attenuation. On the basis of this observation, we set thresholds for the mean attenuation and the total volume of a segmented object: less than –124 HU and greater than 1.5 cm3, respectively. The segmented object had to meet both of these criteria to be considered an ICV. These thresholds were chosen manually from data set 2 so that no true polyps would be lost and then were applied to data set 1.

Computer-aided Polyp Detection Algorithm
After transferring the CT colonographic images to a personal computer (Dell Precision 620 Workstation; Dell, Austin, Tex), a radiologist (R.M.S.) analyzed the images by using our computer-aided polyp detection software package (4,8–12). In brief, voxels along the wall of the colon are identified; then the shape of the wall is determined and the CT attenuation of the wall is measured so that the wall can be classified as polypoid or nonpolypoid. The software analyzed the CT colonographic examinations at a rate of one every 2 minutes; thus, supine and prone examinations were analyzed in a total of 4 minutes.

The polyp detection software used criteria that were developed in earlier studies (13–15). The particular settings for the polyp detector were chosen because they are associated with relatively high sensitivity and a low number of false-positive findings per colon. The finding specifications for the detector were as follows: an elliptical curvature of the peak subtype, a mean curvature range of –2.5 to –0.7 cm–1, eight or more vertexes, a diameter of 0.25 cm or greater, a sphericity of 1.2 or lower, a CT attenuation of the region inside the polyp of between –324 and 976 HU, a minimum segmented polyp volume of 0.1 cm3, a minimum wall thickness of 1 mm, and a minimum polyp surface area of 0.015 cm2. Curvature is a measure of shape, elliptical curvature is an aspect of polypoid shape, vertex and diameter are measures of size, and sphericity is a measure of roundness. Findings that satisfied these criteria were then segmented, as described earlier (7).

The support vector machine classifier is an integral part of the CAD algorithm and has been described previously (15). A genetic algorithm selected the best features for input into the support vector machine (14). Then, a series of support vector machines, each using a different combination of features, were run on detections that passed the filter and the segmenter. If a majority of support vector machines determined that the detected finding represented a polyp, then a positive detection was output by the classifier. This classifier was trained on a third and larger data set that included data set 1 and that was similar to data set 1 in terms of the method of data acquisition and the criteria for patient inclusion. Only detected findings that were segmentable and that passed this classifier were shown to the radiologist as the final CAD output.

A polyp was considered to be detected by the CAD system when the center of a computer-detected polyp was within 1 voxel of the manually drawn contours of a polyp that was recorded in the database and thus matched a polyp described in the colonoscopy report.

Statistical Analyses
P ≤ .05 was considered to indicate significance. The Fisher exact test was used to compare the sensitivities of unmatched data. Confidence intervals for proportions were computed from the binomial distribution.


   Results
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
Results
Discussion
 REFERENCES
 
Twenty-six large polyps from data set 1 and 18 large polyps from data set 2 were retrospectively identified, and the CAD system detected 21 and nine of these polyps, respectively, yielding sensitivities of 81% and 50%, respectively. There were, on average, 4.6 false-positive detections per patient for data set 1 and 2.0 false-positive detections per patient for data set 2.

Data Set 2
For data set 2, 70 (88%) ICVs in 38 patients were detected manually at the 80 supine and prone CT colonographic examinations. Of these 70 ICVs, 11 (16%) that were not segmentable were eliminated by the segmenter. ICVs that were not segmentable did not undergo further processing and were not reported as false-positive polyp detections. Of the 59 segmentable ICVs, 41 (69%) were low in attenuation (less than –124 HU) and 36 (61%) were large (>1.5 cm3). In addition, 25 (42%) of these 59 segmentable ICVs were eliminated by the ICV detector (Fig 1a). Therefore, the total ICV elimination rate was 51% (36 of 70 ICVs).



View larger version (23K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1a. Plots of mean CT attenuation versus volume of true polyps (TP) and ICVs. (a) Data set 2: a training set of 40 patients—20 of whom had normal colonoscopy results and 20 of whom were at high risk for colorectal cancer—with a total of 18 large polyps. Data for the 14 true polyps shown represent those for nine distinct large polyps identified with CAD. The five additional polyps represent multiple detected findings on these nine large polyps and/or findings detected at both supine and prone CT colonography. (b) Data set 1: a test set of 20 patients with 26 known large polyps. The 48 true polyps shown represent 21 distinct large polyps identified with CAD. The 27 additional polyps represent multiple detected findings on these 21 large polyps and/or findings detected at both supine and prone CT colonography. The ICVs are all of those in the data set—regardless of whether they were detected by the CAD system—that were manually identifiable and that could be segmented by the deformable contour segmentation algorithm. True polyps tend to have relatively low attenuation, which is attributable to the partial volume effect. The thick connecting horizontal and vertical lines in a and b show cutoffs for the mean lesion attenuation (less than –124 HU) and the lesion volume (>1.5 cm3) that were found to be optimal for detecting the ICV in this study. Findings in the left upper quadrant of the plot (above and to left of thick lines), which are those of lesions with low attenuation and large volume, are consistent with ICVs and are rejected from the list of polyp candidates reported by the CAD system.

 


View larger version (23K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1b. Plots of mean CT attenuation versus volume of true polyps (TP) and ICVs. (a) Data set 2: a training set of 40 patients—20 of whom had normal colonoscopy results and 20 of whom were at high risk for colorectal cancer—with a total of 18 large polyps. Data for the 14 true polyps shown represent those for nine distinct large polyps identified with CAD. The five additional polyps represent multiple detected findings on these nine large polyps and/or findings detected at both supine and prone CT colonography. (b) Data set 1: a test set of 20 patients with 26 known large polyps. The 48 true polyps shown represent 21 distinct large polyps identified with CAD. The 27 additional polyps represent multiple detected findings on these 21 large polyps and/or findings detected at both supine and prone CT colonography. The ICVs are all of those in the data set—regardless of whether they were detected by the CAD system—that were manually identifiable and that could be segmented by the deformable contour segmentation algorithm. True polyps tend to have relatively low attenuation, which is attributable to the partial volume effect. The thick connecting horizontal and vertical lines in a and b show cutoffs for the mean lesion attenuation (less than –124 HU) and the lesion volume (>1.5 cm3) that were found to be optimal for detecting the ICV in this study. Findings in the left upper quadrant of the plot (above and to left of thick lines), which are those of lesions with low attenuation and large volume, are consistent with ICVs and are rejected from the list of polyp candidates reported by the CAD system.

 
On a per-patient basis, the total elimination rate (ie, with no ICV detection reported on the prone or supine images) was 37% (14 of 38 patients). The mean attenuation and mean volume of the ICVs were –147 HU ± 58 (standard deviation) and 2.3 cm3 ± 1.3, respectively. No true polyps were eliminated. Note that in Figure 1, the true-positive findings tend to have relatively low internal attenuation (<0 HU). This is due to the relatively large partial volume effect that occurs at the edge of polyps.

There were 78 false-positive detections, 18 (23%, in 16 patients) of which were attributed to detection of the ICV. The ICV detection algorithm eliminated 11 (61%, in 10 patients) (95% confidence interval: 36%, 83%) of these 18 false-positive detections. Three cases of eliminated false-positive detections are shown in Figures 2 –4. The seven false-positive detections that were not rejected were associated with seven ICVs and were caused by one or more of the following: segmentation was not successful, the ICV volume was too small, or the CT attenuation of the ICV was too high (Figs 5, 6).



View larger version (92K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2a. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 51-year-old man from data set 2 group. The ICV, which was detected by using CAD and then eliminated by the ICV detector, has a mean attenuation of –131 HU and a volume of 2.7 cm3. The partial fat content of the ICV is evident in a. (c) Marked CT colonographic image shows computer-generated boundary (white outline) of the ICV. Attenuation measurements reflect the mean attenuation within the outline and were obtained in three dimensions, not just on the image shown.

 


View larger version (126K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2b. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 51-year-old man from data set 2 group. The ICV, which was detected by using CAD and then eliminated by the ICV detector, has a mean attenuation of –131 HU and a volume of 2.7 cm3. The partial fat content of the ICV is evident in a. (c) Marked CT colonographic image shows computer-generated boundary (white outline) of the ICV. Attenuation measurements reflect the mean attenuation within the outline and were obtained in three dimensions, not just on the image shown.

 


View larger version (127K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2c. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 51-year-old man from data set 2 group. The ICV, which was detected by using CAD and then eliminated by the ICV detector, has a mean attenuation of –131 HU and a volume of 2.7 cm3. The partial fat content of the ICV is evident in a. (c) Marked CT colonographic image shows computer-generated boundary (white outline) of the ICV. Attenuation measurements reflect the mean attenuation within the outline and were obtained in three dimensions, not just on the image shown.

 


View larger version (90K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3a. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 51-year-old woman from data set 2 group. The ICV, which was detected by the CAD system and then eliminated by the ICV detector, has a mean attenuation of –147 HU and a volume of 2.5 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (131K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3b. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 51-year-old woman from data set 2 group. The ICV, which was detected by the CAD system and then eliminated by the ICV detector, has a mean attenuation of –147 HU and a volume of 2.5 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (140K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3c. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 51-year-old woman from data set 2 group. The ICV, which was detected by the CAD system and then eliminated by the ICV detector, has a mean attenuation of –147 HU and a volume of 2.5 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (127K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 4a. Transverse prone CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 59-year-old man from data set 2 group. The ICV, which was detected by the CAD system and then eliminated by the ICV detector, has a mean attenuation of –155 HU and a volume of 5.0 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (165K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 4b. Transverse prone CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 59-year-old man from data set 2 group. The ICV, which was detected by the CAD system and then eliminated by the ICV detector, has a mean attenuation of –155 HU and a volume of 5.0 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (165K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 4c. Transverse prone CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 59-year-old man from data set 2 group. The ICV, which was detected by the CAD system and then eliminated by the ICV detector, has a mean attenuation of –155 HU and a volume of 5.0 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (111K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 5a. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 53-year-old man from data set 2 group. The ICV, which was detected by the CAD system but not eliminated by the ICV detector because it was too high in attenuation, has a mean attenuation of –90 HU and a volume of 4.9 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (132K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 5b. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 53-year-old man from data set 2 group. The ICV, which was detected by the CAD system but not eliminated by the ICV detector because it was too high in attenuation, has a mean attenuation of –90 HU and a volume of 4.9 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (138K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 5c. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 53-year-old man from data set 2 group. The ICV, which was detected by the CAD system but not eliminated by the ICV detector because it was too high in attenuation, has a mean attenuation of –90 HU and a volume of 4.9 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (114K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 6a. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 56-year-old woman from data set 2 group. The ICV, which was detected by the CAD system but not eliminated by the ICV detector because it was too small, has a mean attenuation of –132 HU and a volume of 1.3 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (138K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 6b. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 56-year-old woman from data set 2 group. The ICV, which was detected by the CAD system but not eliminated by the ICV detector because it was too small, has a mean attenuation of –132 HU and a volume of 1.3 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 


View larger version (151K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 6c. Transverse supine CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 56-year-old woman from data set 2 group. The ICV, which was detected by the CAD system but not eliminated by the ICV detector because it was too small, has a mean attenuation of –132 HU and a volume of 1.3 cm3. (c) Marked CT colonographic image shows the computer-generated boundary (white outline) of the ICV.

 
Data Set 1
For data set 1, 34 (85%) ICVs in 18 patients were detected manually at the 40 supine and prone CT colonographic examinations. The segmenter eliminated 13 (38%) of these 34 ICVs because they could not be segmented. Of the 21 (62%) segmentable ICVs, 10 (48%) were low in attenuation (less than –124 HU) and 13 (62%) were large (>1.5 cm3). The ICV detection algorithm eliminated an additional four (19%) of these 21 ICVs (Fig 1b), for a total elimination rate of 50% (17 of 34 ICVs). On a per-patient basis, the total elimination rate was 28% (five of 18 patients).

With use of parameters that were identical to those used for data set 2, the CAD algorithm made 91 (about four per patient) false-positive detections for data set 1, nine (10%, in seven patients) of which were attributed to detection of the ICV. The ICV detection algorithm eliminated three (33%, in three different patients) of these nine false-positive detections (95% confidence interval: 8%, 70%) and none of the true-positive detections (Fig 7). The six false-positive detections that were not rejected were associated with four ICVs: In two cases, the ICV did not have enough fat content; in one case, the segmenter failed to detect the entire ICV; and in one case, the segmenter mistakenly included part of the terminal ileum.



View larger version (62K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 7a. Transverse prone CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 61-year-old woman from data set 1 group. The ICV, which was detected with CAD and then eliminated by the ICV detector, has a mean attenuation of –140 HU and a volume of 2.0 cm3. (c) Marked CT colonographic image shows computer-generated boundary (white outline) of the ICV.

 


View larger version (129K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 7b. Transverse prone CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 61-year-old woman from data set 1 group. The ICV, which was detected with CAD and then eliminated by the ICV detector, has a mean attenuation of –140 HU and a volume of 2.0 cm3. (c) Marked CT colonographic image shows computer-generated boundary (white outline) of the ICV.

 


View larger version (131K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 7c. Transverse prone CT colonographic images obtained by using (a) soft-tissue and (b, c) lung window settings show ICV (arrow) in 61-year-old woman from data set 1 group. The ICV, which was detected with CAD and then eliminated by the ICV detector, has a mean attenuation of –140 HU and a volume of 2.0 cm3. (c) Marked CT colonographic image shows computer-generated boundary (white outline) of the ICV.

 
The performance of the ICV detector on manually detected ICVs and on false-positive polyp findings was not significantly different between data sets 1 and 2 with use of either per-ICV or per-patient elimination rates (P ≥ .34).


   Discussion
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
Results
 Discussion
REFERENCES
 
We developed a CAD algorithm that can identify many ICVs and that does not mistake a true polyp for the ICV. The algorithm consists of a two-stage procedure: (a) A segmenter seeks convex segmentable objects and a pair of thresholds based on volume and CT attenuation. (b) The algorithm produces a benefit by reducing the number of false-positive detections of the ICV made with CAD: In this study, there was a 33% reduction in false-positive detections for data set 1 and a 61% reduction in false-positive detections for data set 2.

Each false-positive detection requires inspection by the radiologist, increasing interpretation time and examination costs. In addition, if the ICV is routinely detected by a CAD program, the program becomes ineffective in identifying lesions on or near the ICV. This is because the radiologist may assume that any detected finding in the cecum most likely represents a false-positive finding on the ICV. Although it is desirable to have no false-positive detections per patient, this is not feasible in practice. Polypoid structures can appear in the colon by chance—for example, because of extrinsic compression of the colon by a structure, such as the liver or another bowel loop. If the colon is inadequately distended, haustral folds can appear thick and polypoid, leading to false-positive detections with CAD.

In addition, as CAD algorithms improve and false-positive rates decrease, the gain in interpretation efficiency will proportionately increase with use of a method such as ours that rejects normal polypoid structures such as the ICV. In other words, the added benefit of reducing the number of false-positive detections from one to zero will be greater than a decrease from four to three false-positive detections as a fraction of the total interpretation time.

The ICV is a polypoid structure that can mimic a colonic mass. In our study, the ICV was frequently visible: at 104 (87%) of 120 examinations. This experience is similar to that reported in a study in which barium enema examinations were performed: The ICV was visible in 91 (86%) of 106 patients (16). In our experience, the ICV had a variety of shapes. At barium enema examination, the ICV is usually smooth (in 85% of cases) or smoothly lobulated (in 15% of cases), and it may be large or asymmetric, even in the absence of tumorous tissue (16). Even the relationship between the ICV and the cecal haustral folds can be variable, with the majority of ICVs located on either the first or the second fold (16). In addition, the ICV can be located medially (in 81% of cases), laterally (in 15% of cases), or posteriorly (in 3% of cases) (16).

There is little information about the imaging characteristics of the normal ICV at CT in the literature. The ICV often contains fat, or lipomatous infiltration (17). Barium enema examinations do not yield data about the internal composition of the ICV. We found that the majority (51 [64%] of 80) of segmentable ICVs had relatively low internal attenuation, which is consistent with fat.

There are different ways to consider the sensitivity of the described ICV detector. One way is to consider all ICVs and how many were accurately recognized by the detector. Another way is consider only the ICVs that were detected by the CAD system and how many of these were eliminated. Both methods are valid because the CAD system can change and different ICVs may be detected.

Our ICV detector identified ICVs well in both data sets, despite differences in image acquisition techniques (single– vs multi–detector row), scanning preparations (with vs without administration of glucagon), and indications for scanning (known polyps vs surveillance of high-risk population). This result indicates that the depiction of the ICV was not affected to a great extent by these differences and that the algorithm is relatively robust. Because the ICV detector was trained by using data set 2 and tested on data set 1, it is not surprising that its sensitivity was greater for data set 2, both on manually detected ICVs and on false-positive polyp findings, although the differences were not statistically significant. The lack of a significant difference in the fractional reduction of false-positive polyp findings (P = .34) between the two data sets may also be attributable in part to the relatively small sample size. In addition, data set 1 was obtained by using a single–detector row helical CT scanner. This factor could also explain why performance was lower for data set 1.

We used 5-mm collimation with a 2-mm overlap in this study to maintain consistency with our previous research parameters (4). We have observed that polyps have higher attenuation if the collimation is thinner, because the partial volume effect is reduced. If thinner collimation is used, the thresholds will probably need to be readjusted. The effect of section thickness on the accuracy of ICV attenuation measurement will need to be determined.

We did not report results for polyps smaller than 1.0 cm because our CAD system was not trained to identify them. In addition, polyps smaller than 1.0 cm were more difficult to identify in retrospect. With use of thinner collimation, we may be able to train a CAD system to detect smaller polyps. Nevertheless, because the ICV detector was trained to recognize structures with volumes larger than 1.5 cm3, it is unlikely that it would inappropriately exclude small polyps. In addition, our ICV detector would probably eliminate many lipomas. Although lipomas are abnormal growths, they have no malignant potential, and, thus, it is much less important that they be detected.

We did not use location information to analyze true polyps and false-positive findings according to whether they were located in the cecum. We assumed that polyps located elsewhere in the colon were similar in CT appearance to those in the cecum, and we used the other polyps as surrogates for cecal polyps, which must be distinguished from ICVs. It is said that right-sided colonic polyps have different genetic compositions compared with polyps located elsewhere in the colon (18,19). Whether right-sided and other colonic polyps have different shapes and CT attenuations is still unknown. Although it is possible that the performance of the ICV detector could be improved with use of location information, our data on the true polyps are not sufficient to conclude that the use of this information would lead to any substantial performance enhancement. In clinical practice, the ICV detector would classify a false-positive finding as the ICV only if the finding was located in the cecum.

There is a potential pitfall in the setting of abnormal ICVs. Polyps and masses can occur on the ICV, although this is uncommon. Because of a lack of adequate data, we did not test the behavior of our algorithm with abnormal ICVs. However, real polyps have the attenuation of soft tissue—not fat—even if they are located on the ICV; this factor could serve to distinguish them from normal ICVs.

The CAD software had relatively low sensitivity, particularly for data set 2. Although this may have been due in part to the relatively thick section collimation, note that the patient population was a surveillance group of individuals at high risk for colorectal cancer who had undergone colonoscopy during the past few years. These patients may have had polyps that were missed at colonoscopy or that had grown in the interval since they underwent colonoscopy. It is unknown whether such polyps have different geometric characteristics that might make them more difficult to detect. Indeed, the radiologists had difficulty detecting these polyps, and the CAD software identified some polyps that were missed by the radiologists (5). In summary, in our study, we showed how a CAD algorithm can help to reduce the frequency of false-positive detections related to ICVs.


   ACKNOWLEDGMENTS
 
We thank Andrew Dwyer, MD, of the National Institutes of Health, for critical review of the manuscript.


   FOOTNOTES
 
Abbreviations: CAD = computer-aided detection, ICV = ileocecal valve

The authors have pending and/or awarded patents for the subject matter described in the manuscript. C.D.J. has software licensed to GE Medical Systems.

Author contributions: Guarantor of integrity of entire study, R.M.S.; study concepts, R.M.S., C.D.J.; study design, R.M.S., J.Y.; literature research, R.M.S.; clinical studies, C.D.J.; experimental studies, R.M.S.; data acquisition, C.D.J.; data analysis/interpretation, R.M.S., J.Y.; statistical analysis, R.M.S., J.Y.; manuscript preparation, R.M.S.; manuscript definition of intellectual content, revision/review, and final version approval, R.M.S., J.Y., C.D.J.; manuscript editing, R.M.S., J.Y.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
Results
 Discussion
 REFERENCES
 

  1. Summers RM, Yoshida H. Future directions: computer-aided diagnosis. In: Dachman AH, eds. Atlas of virtual colonoscopy. New York, NY: Springer, 2003; 55-62.
  2. Summers RM. Challenges for computer-aided diagnosis for CT colonography. Abdom Imaging 2002; 27:268-274.[Medline]
  3. Dachman AH, Nappi JJ, Frimmel H, Yoshida H. Sources of false positives in computerized detection of polyps in CT colonography (abstr). Radiology 2002; 225(P):303-304.
  4. Summers RM, Johnson CD, Pusanik LM, Malley JD, Youssef AM, Reed JE. Automated polyp detection at CT colonography: feasibility assessment in a human population. Radiology 2001; 219:51-59.[Abstract/Free Full Text]
  5. Summers RM, Jerebko A, Franaszek M, Malley JD, Johnson CD. Colonic polyps: complementary role of computer-aided detection in CT colonography. Radiology 2002; 225:391-399.[Abstract/Free Full Text]
  6. Hara AK, Johnson CD, Reed JE, et al. Reducing data size and radiation dose for CT colonography. AJR Am J Roentgenol 1997; 168:1181-1184.[Free Full Text]
  7. Yao J, Miller M, Summers RM. Automatic segmentation and detection of colonic polyps in CT colonography based on knowledge-guided deformable models. In: Clough AV, Amini A, eds. SPIE medical imaging. San Diego, Calif: International Society for Optical Engineering, 2003; 5031:370-380.
  8. Summers RM, Feng DH, Holland SM, Sneller MC, Shelhamer JH. Virtual bronchoscopy: segmentation method for real-time display. Radiology 1996; 200:857-862.[Abstract/Free Full Text]
  9. Summers RM. Navigational aids for real-time virtual bronchoscopy. AJR Am J Roentgenol 1997; 168:1165-1170.[Abstract/Free Full Text]
  10. Summers RM. Image gallery: a tool for rapid endobronchial lesion detection and display using virtual bronchoscopy. J Digit Imaging 1998; 11:53-55.
  11. Summers RM, Beaulieu CF, Pusanik LM, et al. Automated polyp detector for CT colonography: feasibility study. Radiology 2000; 216:284-290.[Abstract/Free Full Text]
  12. Summers RM, Jerebko AK, Franaszek M, Malley JD. An integrated system for computer-aided diagnosis in CT colonography: work-in-progress. In: Lemke HU, Vannier MW, Inamura K, Farman AG, Doi K, eds. Computer assisted radiology and surgery (CARS). Presented at the 15th International Congress and Exhibition, Berlin, Germany, June 27–30. 2001.
  13. Jerebko AK, Malley JD, Franaszek M, Summers RM. Multiple neural network classification scheme for detection of colonic polyps in CT colonography data sets. Acad Radiol 2003; 10:154-160.[CrossRef][Medline]
  14. Miller M, Jerebko AK, Malley JD, Summers RM. Comparison of stochastic vs heuristic feature subset selection methods for improving classification algorithms in computer-aided detection. In: Clough AV, Amini A, eds. SPIE medical imaging. San Diego, Calif: International Society for Optical Engineering, 2003; 5031:102-110.
  15. Malley JD, Jerebko AK, Summers RM. Committee of support vector machines for detection of colonic polyps from CT scans. In: Clough AV, Amini A, eds. SPIE medical imaging. San Diego, Calif: International Society for Optical Engineering, 2003; 5031:570-578.
  16. El-Amin LC, Levine MS, Rubesin SE, Shah JN, Kochman ML, Laufer I. Ileocecal valve: spectrum of normal findings at double-contrast barium enema examination. Radiology 2003; 227:52-58.[Abstract/Free Full Text]
  17. Fultz PJ, Hampton WR, Skucas J, Sickel JZ. Differential diagnosis of fat-containing lesions with abdominal and pelvic CT. RadioGraphics 1993; 13:1265-1280.[Abstract/Free Full Text]
  18. Okamoto T, Konishi F, Kojima M, Senba S, Kanazawa K, Tsukamoto T. Significance of microsatellite instability in different types of early-stage nonfamilial colorectal carcinomas. Dis Colon Rectum 1998; 41:1385-1391.[CrossRef][Medline]
  19. Thibodeau SN, Bren G, Schaid D. Microsatellite instability in cancer of the proximal colon. Science 1993; 260:816-819.[Abstract/Free Full Text]



This article has been cited by other articles:


Home page
 Am. J. Roentgenol.Home page
K. T. Johnson, J. G. Fletcher, and C. D. Johnson
Computer-Aided Detection (CAD) Using 360{degrees} Virtual Dissection: Can CAD in a First Reviewer Paradigm Be a Reliable Substitute for Primary 2D or 3D Search?
Am. J. Roentgenol., October 1, 2007; 189(4): W172 - W176.
[Abstract] [Full Text] [PDF]

Home page
 Am. J. Roentgenol.Home page
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]

Home page
 RadioGraphicsHome page
A. C. Silva, S. D. Beaty, A. K. Hara, J. G. Fletcher, J. L. Fidler, C. O. Menias, and C. D. Johnson
Spectrum of Normal and Abnormal CT Appearances of the Ileocecal Valve and Cecum with Endoscopic and Surgical Correlation
RadioGraphics, July 1, 2007; 27(4): 1039 - 1054.
[Abstract] [Full Text] [PDF]

Home page
 RadiologyHome page
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]

This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Summers, R. M.
Right arrow Articles by Johnson, C. D.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Summers, R. M.
Right arrow Articles by Johnson, C. D.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
RADIOLOGY RADIOGRAPHICS RSNA JOURNALS ONLINE