| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Original Research |
1 Department of Radiology, Mayo Clinic, Rochester, MN 55905.
2 Present address: Department of Radiology, Mayo Clinic, 13400 E Shea Blvd.,
Scottsdale, AZ 85259.
Received October 17, 2006;
accepted after revision May 3, 2007.
J. G. Fletcher received an educational license from GE Healthcare, the
developer of the software being studied. The monies from that educational
license are being used to support his research program. He has not received
personal compensation from GE Healthcare and has no ownership or royalty
arrangement with GE Healthcare with respect to the virtual dissection and
computer-aided detection software evaluated in this study.
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
MATERIALS AND METHODS. Forty-one consecutive patients who underwent imaging without oral contrast material for stool tagging from a teaching file database constituted the patient population for this feasibility study. Using CT colonography equipped with CAD software, reviewers evaluated each possible polyp detected by the software using virtual dissection images combined with axial and 3D endoluminal views and compared the results with optical colonoscopy, the reference standard. Two experienced radiologists blinded to the reference standard findings interpreted the CAD detections to be true or false. The false detections were reviewed and categorized.
RESULTS. Sensitivities for polyps that were 6–9 mm were 78.3%
(18/23) and 91.3% (21/23) for reviewers 1 and 2, respectively. For polyps 
1 cm, sensitivities were 94.9% (37/39) and 97.4% (38/39) for reviewers 1 and
2, respectively. Per-patient sensitivities for polyps 6 and 10 mm
were 94.4% (34/36) and 95.1% (39/41) for reviewer 1 and 97.2% (35/36) and
97.6% (40/41) for reviewer 2, respectively. The average number of false
detections per acquisition was 4.28. The average interpretation times were 4
minutes 26 seconds and 5 minutes 38 seconds for reviewers 1 and 2,
respectively.
CONCLUSION. Colorectal polyp detection using CT colonography equipped with CAD and virtual dissection as a first reviewer is feasible. Detection rates are similar to colonoscopy. Interobserver variability is low and interpretation times are short. False-positive detections per patient are few in number.
Keywords: CAD • colorectal cancer • computer-aided detection • CT colonography • virtual dissection
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Even if lesions are detected by the CAD software, radiologists must characterize the detected lesions as polyps or tumors versus common sources of false-positive findings such as folds, stool, and the ileocecal valve. Summers et al. [4] found that CAD does help to increase sensitivity at CT when used as a second reviewer by detecting polyps that had been missed by radiologists. The reported sensitivity for large polyps using CAD as a second reviewer was 89%, with an average of 11 false-positives per patient [4]. Another large patient study conducted by Summers et al. [3] using an improved CAD program found the sensitivity for large polyp detections using CAD as a second reviewer to be equally high, but the false-positive detection rate was reduced to 2.1 per patient. Mani et al. [6] reported a maximum sensitivity for colorectal lesion detection of 74% using CAD as a first reviewer. To our knowledge, no reports exist about using CAD as a first reviewer using virtual dissection images.
The purpose of this study was to evaluate the feasibility of a new CAD software program as a first reviewer for detecting colorectal polyps when applied to 360° virtual dissection image display.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
All patients underwent CT colonography before same-day colonoscopy. Preparation for the examination included an electrolyte lavage (GoLyte, Braintree Laboratories) and bisacodyl tablets (10 mg); or magnesium citrate (300 mL) and bisacodyl tablets (20 mg) or sodium phosphates solution (Phospho-Soda, Fleet) (90 mL). Glucagon (1 mg) was given subcutaneously 10 minutes before CT acquisition unless contraindicated or refused by the patient. Patients were placed in the left lateral decubitus position for enema tip insertion and slow manual insufflation of approximately 2 L of carbon dioxide (or until the patient verbally indicated maximal tolerance).
Before each acquisition, a CT scout image was obtained to confirm complete colon filling; both supine and prone data acquisitions were performed. Additional carbon dioxide was added as tolerated by the patient as needed to fully distend the colon. All of the examinations were performed using an 8-MDCT scanner (LightSpeed Ultra, GE Healthcare). After colon insufflation, a breath-hold anteroposterior scout image was obtained to assess luminal distention and to prescribe axial slices through the entire large bowel. Images were acquired using 2.5-mm collimation; table speed, 13.5 mm/s (pitch of 1.35); 1.25-mm reconstruction intervals; matrix, 512 x 512; field of view, to fit; 120–240 mA (120 mA was doubled if patient's transverse diameter was > 40 cm); tube rotation time, 0.5 second; 120 kVp; standard reconstruction algorithm; and 28-second breath-holds.
Colonoscopy was performed in the standard fashion to examine the entire colorectum. All of the examinations were performed by staff gastroenterologists or colorectal surgeons experienced with this technique. All polyps were proven at colonoscopy. Lesion size was determined using the pathology report unless the lesion had been removed in pieces. In those cases, the estimate of lesion size at colonoscopy was used.
The 41 patient data sets were evaluated on a workstation (AW [version 4.2, 2005], GE Healthcare) using CT colonography software (Vox-tool, version 7.1.38, GE Healthcare). A midline trace for both the supine and prone data sets was automatically (n = 37) or semiautomatically (n = 45) created through the lumen of the colon, with human confirmation and assistance if required. The program displays interactive and synchronized supine and prone axial, 3D endoluminal, and virtual dissection images on two screens. The virtual dissection image depicts 360° of the colon lumen. The CAD software is based on a proprietary shape-based recognition algorithm. Detected lesions are highlighted on their surface on all three image displays.
In a blinded fashion, two experienced radiologists (> 1,000 proven cases at CT colonography and > 50 cases with virtual dissection) interpreted the virtual dissection images by assessing only CAD detections to be true polyps or false detections. These cases had not been reviewed by either radiologist during the prior 12 weeks. Only lesions detected by the CAD software were reviewed. The axial and 3D endoluminal views were used as needed for problem solving when reviewing CAD detections and to improve reviewer confidence. After the images were reviewed and data were collected, the colonoscopy findings were revealed and the cause of each false detection was noted and categorized as ileocecal valve, enema tip, stool, fold, or other. Interpretation times were recorded excluding the loading and tracing times (time to create midline trace).
The CAD algorithm is a 3D filtering algorithm that uses local curvature at
implicit isosurfaces to identify shapes. The algorithm is designed to filter
the data for specific shapes from a model-based approach and does not require
a training database. This method, termed "curvature tensor,"
determines the minimum (kmin) and maximum
(kmax) local curvatures in the null space of the gradient
in a 3D volume instead of a surface. The respective curvatures can be
determined using the following formula (equation 1):
 | (1) |
 | (2) |
The responses of the curvature tensor (kmin and kmax) are segregated into spherical and cylindric responses on the basis of thresholds on kmin, kmax, and the ratio of kmin/kmax derived from the size and aspect ratio of the sphericalness and cylindricalness. Masking out regions that have a cylindric-shaped overlap with the spherical responses performs a cleanup of erroneous responses. The resulting spherical regions are displayed in 2D, 3D endoluminal, and virtual dissected views as highlighted overlays.
Sensitivity was determined by assessing the number of CAD detections that were assessed by the radiologist to be real polyps divided by the total number of polyps as determined by the reference standard (colonoscopy).
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
5
mm, 23 (24.2%) were 6–9 mm (Fig.
1A,
1B,
1C), and 37 (38.9%) polyps were
1 cm (Fig. 2A,
2B). Of the 95 polyps, 12
(12.6%) were carcinomas (Fig.
3) and eight polyps (8.4%) were flat, 10 (10.5%) were
pedunculated, and 65 (68.4%) were sessile in morphology. Eleven (11.6%) of the
95 polyps were located in the cecum, 15 (15.8%) in the ascending colon, 13
(13.7%) in the transverse colon, eight (8.4%) in the descending colon, 17
(17.9%) in the sigmoid colon, 18 (18.9%) in the rectum, and 11 (11.6%) in
either the splenic or hepatic flexure; the remaining two polyps (2.1%) were
cancers located in the appendix region that had spread into the colon. The
midline trace used as part of the software package required human interaction
and correction in 45 cases (47.4%) and was entirely satisfactory and automatic
in 37 (38.9%).
|
|
|
|
|
|
1 cm, the sensitivity ranged between 94.9% (37/39) and 97.4%
(38/39) for reviewers 1 and 2, respectively
(Table 1). Per-patient
sensitivities for detecting colorectal polyps 6 and 10 mm were 94.4%
(34/36) and 92.7% (38/41) for reviewer 1 and 97.2% (35/36) and 97.6% (40/41)
for reviewer 2, respectively (Table
1). Reviewer 1 also misinterpreted two additional polyps, one in
the ascending colon and one in the rectum, that were flat or pedunculated
(Fig. 4A,
4B,
4C).
|
|
|
|
The average number of false detections (CAD detections that were not polyps) per patient was 4.28 (range, 1–15). The causes of the false detections for reviewer 1 included the enema tip (n = 36) and ileocecal valve (n =40) on both the supine and prone images. The number of false detections that were classified as stool or folds was 128 and 101, respectively, for reviewer 1 (Table 2). Both reviewers called false detections as polyps in 10 (24%) of the 41 patients. Interpretation times ranged from 1 minute 12 seconds to 11 minutes 13 seconds for reviewer 1, with an average of 4 minutes 26 seconds per case, and from 1 minute 19 seconds to 14 minutes for reviewer 2, with an average of 5 minutes 38 seconds per case.
|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The sensitivity for polyp detection remained high despite the fact that the radiologists evaluated only CAD detections. Interobserver variability was markedly reduced while the high sensitivity was preserved. For both reviewers, the sensitivity using CAD was similar to the reported sensitivity at colonoscopy for the detection of large lesions. False detections were relatively few, averaging 4.28 per patient. This likely accounts for the short interpretation times. Stool tagging, which was not used in this study, combined with stool subtraction should make discrimination of stool from polyps easier in the future.
An important finding was that CAD software with virtual dissection revealed seven (88%) of eight flat lesions. Only one of eight polyps was not identified by CAD. This high detection rate could indicate that CAD is a useful adjunct for detecting these difficult-to-find lesions because flat lesions are frequently missed by unassisted human observation [8].
The ileocecal valve and the enema tip were common false detections, occurring in almost half of the data sets. Fortunately, these cases of false detection are easy to recognize. More advanced software programs should be able to eliminate these common causes of false detections in the future.
CAD dramatically reduced interobserver variability. In this study, sensitivities ranged between 94.9% and 97.4% for reviewers 1 and 2, respectively. This variability is smaller than that previously found to occur at CT colonography. Johnson et al. [9] reported that reviewer sensitivities varied from 32% to 73%. CAD reduces the variability of polyp detection by automated detections that should be the same for all reviewers. In this study, interobserver variability was due to differences in characterizing detections as real or not.
Although the performance of CAD in this study is high, it did not detect all the lesions that were present. Further study of its performance in a low-prevalence screening population is needed.
There were several limitations to this study. First, two patients were excluded from the study because oral contrast material was used for colon preparation. This CAD program did not use a tagged-stool subtraction algorithm. If these cases had been evaluated, the number of false detections would have been unreasonably high. In addition, only the polyp candidates detected by the CAD software were eligible as true polyps. Even if the radiologist found a polyp that was not detected by the CAD software, that polyp was not considered in the data analysis. Therefore, sensitivity estimates may be even higher when both the CAD software and the radiologist detections are combined.
The population examined had a very high incidence of colorectal neoplasms and is not representative of a screening population. This study, therefore, serves as a feasibility study using CAD as a first reviewer. Further study using this methodology in a screening population is warranted. Estimates of specificity were not possible because all patients had polyps. This issue is an important one to evaluate in the future because low specificity could increase the real costs of this examination.
The use of cases from a teaching file represents a selection bias for the types of cases included (i.e., polyps or cancers that are not occult) [10]. On the other hand, as teaching cases, many represented difficult-to-detect lesions that are commonly overlooked. Flat lesions were overrepresented in our study group—compared with a screening population—because of the need to emphasize these lesions in teaching and as evidence of the challenging nature of this study set.
A number of reports on about the performance of CAD systems have been
published recently; some evaluate systems for polyp detection, and others test
how these systems are used by humans
[5,
6,
11–13].
Summers et al. [5] has the
largest series (792 patients with 29 polyps) with an overall sensitivity for
polyps 1 cm of 89% and a mean of 2.1 false-positive detections per
patient. In a smaller series, other authors have reported similar performance,
although the number of false-positives differs. Halligan et al.
[11] reported 10–32
false-positives depending on sensitivity for polyp detection. Our results are
associated with some of the highest reported sensitivity rates with an
intermediate number of mean false detections using the novel virtual
dissection display.
In conclusion, colorectal polyp detection using CAD and virtual dissection as a first reviewer is feasible. Detection rates are similar to colonoscopy. Interobserver variability is low and interpretation times are short. False-positive detections per patient are few in number. A larger trial in a screening population is warranted to further evaluate this promising technology.
Acknowledgments
We thank Debora Shreve for her help in preparing this manuscript. We also
thank Saad Sirohey for the description of how the CAD system described detects
polyp candidates. The information supplied by S. Sirohey is included in the
Introduction with his permission.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  Progress in Shortening Reading Times for CT Colonography Journal Watch Gastroenterology, November 2, 2007; 2007(1102): 1 - 1. [Full Text]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
