| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Gastrointestinal Imaging |
1 From the Department of Radiology, Stanford University Medical Center, James H. Clark Center, 318 Campus Dr, Room S324, Stanford, CA 94305-5450 (R.S., P. Schraedley-Desmond, S.N., E.W.O., R.B.J., D.M., D.S.P., A.J.S., P. Sundaram, C.F.B.); Veterans Affairs Palo Alto Health Care System, Palo Alto, Calif (E.W.O.); Department of Radiology, University of California, San Francisco, San Francisco, Calif (J.Y.); San Francisco Veterans Affairs Medical Center, San Francisco, Calif (J.Y.); and Department of Radiology, Massachusetts General Hospital, Boston, Mass (M.E.Z.). Received March 11, 2005; revision requested May 4; revision received May 20; accepted June 20; final version accepted August 24. Supported by the National Institutes of Health (R01 CA72023 and 1 U54 GM072970) and the Lucas Foundation. Address correspondence to R.S. (e-mail: rshi{at}stanford.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
Materials and Methods: Institutional review board approval and informed consent were obtained; this study was HIPAA compliant. Forty-seven computed tomographic (CT) colonography data sets were obtained in 26 men and 10 women (age range, 42–76 years). Four radiologists classified 705 polyp candidates (53 TP candidates, 652 FP candidates) identified with CAD; initially, only two-dimensional images were used, but these were later supplemented with 3D rendering. Another radiologist unblinded to colonoscopy findings characterized the features of each candidate, assessed colon distention and preparation, and defined the true nature of FP candidates. Receiver operating characteristic curves were used to compare readers' performance, and repeated-measures analysis of variance was used to test features that affect interpretation.
Results: Use of 3D viewing improved classification accuracy for three readers and increased the area under the receiver operating characteristic curve to 0.96–0.97 (P < .001). For TP candidates, maximum polyp width (P = .038), polyp height (P = .019), and preparation (P = .004) significantly affected accuracy. For FP candidates, colonic segment (P = .007), attenuation (P < .001), surface smoothness (P < .001), distention (P = .034), preparation (P < .001), and true nature of candidate lesions (P < .001) significantly affected accuracy.
Conclusion: Use of 3D viewing increases reader accuracy in the classification of polyp candidates identified with CAD. Polyp size and examination quality are significantly associated with accuracy.
© RSNA, 2006
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
Thirty-six subjects were selected for the present study (26 men; 10 women; age range, 42–76 years; mean age, 59 years) between November 2001, when both institutions started to use multidetector CT for CT colonography, and March 2004. Fourteen subjects were selected from Stanford University Medical Center, and 22 were selected from San Francisco Veterans Affairs Medical Center. Twenty-six consecutive subjects were chosen because (a) the presence of at least one polyp larger than 5 mm was confirmed with same-day colonoscopy, (b) preparation was adequate (ie, complete colonoscopy examination, without limited viewing because of retained stool), and (c) no segmental collapse of the colon was detected. The other 10 subjects were randomly chosen from a group of patients in whom no polyps were detected with colonoscopy during the same time period.
Twenty-two subjects underwent bowel cleansing with CoLyte (Schwarz Pharma, Milwaukee, Wis) and Citrate of Magnesia (Humco, Texarkana, Tex); 14 subjects received GoLytely (Braintree Laboratories, Braintree, Mass) on a schedule similar to that used by Macari et al (18).
CT Colonography and Optical Colonoscopy
The colon was insufflated with room air and a manual bulb according to subject tolerance. Each subject was examined in the supine and prone positions with a multidetector CT scanner: Patients were examined with either a LightSpeed 16 scanner (GE Healthcare, Milwaukee, Wis) (n = 14), a LightSpeed Ultra scanner (GE Healthcare) (n = 18), or a LightSpeed Plus scanner (GE Healthcare) (n = 4). Examinations were performed with 120-kV tube voltage, 50–200-mA tube current, and 0.5-second gantry rotation. All images were reconstructed by using a standard kernel with 2.5-mm section thickness and 1.25-mm reconstruction intervals.
All subjects underwent optical colonoscopy within 1–2 hours of CT colonography, with use of standard practices. Polyp location and size were recorded in a database. Endoscopic size estimates (obtained with open biopsy forceps) were used to compare optical colonoscopy measurements with CT colonography measurements. Enumeration of polyp count and size range is reported in the CT colonography reference standard section.
CAD Algorithm
We used a previously validated CAD algorithm known as the surface normal overlap method (14). The output of this algorithm is a list of candidate polyps and their voxel coordinates, where each candidate is ranked in descending order according to the number of intersecting or nearly intersecting surface normal vectors. Each entry in a given CAD list could represent either a TP or an FP finding.
Reader Interface
Both the pretrial characterization of TP and FP polyp candidates and the blinded readings were performed with a locally developed personal computer–based platform (Stanford University Radiology Framework [SURF]) (Fig 1) (19). In this study, we used a personal computer with a 3.0-GHz Pentium IV processor (Intel, Santa Clara, Calif), 2 GB of random access memory, and a 17-inch (43.18-cm) color cathode-ray tube display. Interactive 3D texture-based volume rendering with the SURF platform was computed with a graphics card (NVIDIA GeForce4 Ti4600; Nvidia, Santa Clara, Calif) by using opacity and color tables similar to those described by McFarland et al (20). The top two windows of the interface displayed transverse or coronal images and permitted interactive panning and zooming, size measurement, and window and level setting control (used only to set reference standard). A third window, which could be toggled on or off with a keystroke, displayed an interactive 3D volume rendering of a cubical subvolume (32 pixels on each side) centered on a selected polyp candidate. A fourth text-based window, referred to as the polyp CAD manager (Fig 2), allowed the reader to select from the list of polyp candidates. During blinded reading, readers could enter their confidence level for each candidate. Subpanels demonstrating reference standard information for each polyp candidate were not available during blinded reader evaluations.
|
|
In the first phase, full interpretation of all optical colonoscopy studies with positive findings was performed by viewing images obtained in the supine and prone positions together on a commercial 3D workstation (Vitrea 2; Vital Images, Plymouth, Minn). A two-dimensional (2D) search and a 3D problem-solving approach (21) were used to localize and mark all polyps. To be considered a TP polyp, a lesion had to be located in the same colon segment and have a maximum width within 50% of the diameter of the optical colonoscopy finding. A text file of TP polyps was saved and imported to the SURF platform for the second stage of reference standard setting.
Not all polyps were depicted on both supine and prone images (22). On the basis of polyp visibility to the unblinded radiologist, we included data obtained in 11 patients in the supine and prone positions. To provide unique polyps to the readers, however, images obtained in only one position were included for each polyp. Polypoid structures identified with CT colonography but not optical colonoscopy were interpreted as FP findings. Polypoid structures identified by the unblinded CT colonography reader but not detected with CAD or optical colonoscopy were excluded. The second phase of reference standard setting consisted of generating lists of polyp candidates and systematic characterization of all of the candidates.
At its current stage of development, CAD is relatively limited in that most algorithms perform poorly in the evaluation of small (<10 mm) polyps. We chose an arbitrary number of 15 candidate lesions and simulated a detector that would include all TP polyps, regardless of size, in the top 15 output. We first included all TP polyps found during the first phase. To complete a list of 15 candidates, we then added the FP polyp candidates with highest scores and randomized each list.
At the end of this process, a positive data set resulted in a candidate list containing between one and four TP polyps and enough FP findings to yield a list of 15 candidates. Control data sets contained 15 FP findings. While the process of generating reading lists was somewhat contrived, this study was not intended to evaluate CAD performance but to present a reasonable number of candidate lesions to the blinded readers to evaluate. In other words, regardless of the construction of the reading list, the variables tested were differences between 2D viewing and combined 2D and 3D viewing and imaging features of each polyp candidate.
For all polyp candidates, the unblinded radiologist used a five-point scale to assign a confidence rating: 1, definitely not a polyp; 2, probably not a polyp; 3, equivocal for polyp; 4, probably a polyp; and 5, definitely a polyp.
Next, the unblinded radiologist characterized a set of additional features for each polyp candidate with the polyp CAD manager by assigning values to each of the fields shown in Figure 2. Features characterized included size, shape, attenuation, and surface smoothness. Three size characteristics (maximum width, minimum width, and height [defined as vertical height above the surface]) were measured on transverse and coronal images. Shape and attenuation were categorical variables: Shape was characterized as pedunculated, sessile, flat, or not applicable. Attenuation was characterized as homogeneous or heterogeneous if visible air or fat was contained. Surface smoothness was considered a continuous variable and scored with a four-point scale: 1, very irregular; 2, somewhat irregular; 3, mostly smooth but partially irregular; and 4, completely smooth.
Features relating to the colon itself near each polyp candidate included colonic segment in which the candidate was located, binary assessment of location of the candidate on a haustral fold, assessment of local distention, and quality of preparation. The colon was divided into eight segments (ie, cecum, ascending colon, hepatic flexure, transverse colon, splenic flexure, descending colon, sigmoid colon, and rectum). Distention was graded with a four-point scale: 1, collapsed; 2, 25%–50% distended; 3, 51%–75% distended; and 4, 76%–100% distended. Preparation quality was also graded with a four-point scale: 1, more than 50% of the lumen filled with retained fluid, solid material, or both; 2, 25%–50% of the lumen filled with retained material; 3, less than 25% of the lumen filled with retained material; and 4, no retained material.
The unblinded radiologist identified FP findings (ie, stool, cancer, lipoma, thickened haustral fold, diverticulum, imaging artifact, ileocecal valve, or small bowel). We considered cancer (no instances in our data) and lipoma (one instance) to be FP findings, as their appearances are distinct from the appearance of the TP polyps (including both adenomatous and hyperplastic polyps) we wished to assess. Rarely, small-bowel FP polyp candidates were detected because of imperfect segmentation.
There were 47 data sets presented; 37 contained a polyp or polyps, and 10 did not. Each data set comprised 15 experimental polyp candidates, yielding 705 polyp candidates (53 TP candidates and 652 FP candidates). These included six polyps 10 mm in diameter or larger in five data sets, 23 polyps between 5.0 and 9.9 mm in diameter in 20 data sets, and 24 polyps smaller than 5.0 mm in diameter in 18 data sets. Note that one positive data set could contain more than one polyp, possibly in a different size range.
Blinded Reader Trial
Four radiologists (E.W.O., R.B.J., J.Y., M.E.Z.) served as independent readers. Each had extensive experience with CT colonography research (3–9 years of experience; average experience, 6.5 years), had interpreted 100–1000 CT colonography studies, and had participated in CT colonography reader trials. Two readers were unfamiliar with the SURF platform, whereas two others had participated in a pilot trial in which 10 different data sets were used. The study coordinator orientated each reader to the SURF platform and the widgets; three data sets containing 30 polyp candidates were used in the orientation session. Readers were informed that all experimental polyp candidate lists could contain both TP and FP polyp candidates, as well as normal structures (ie, the ileocecal valve). Training data sets were not included in the blinded trial. The study coordinator informed readers of the study goals but did not reveal the prevalence or size range of TP lesions. Readers were informed that the prevalence of polyps was higher in this study than in a typical screening cohort.
Data sets were viewed in random order by each of the readers during two to four sessions, each lasting 2–4 hours. Data sets were first loaded for 2D viewing with window and level settings of 1500 HU and –500 HU, respectively. Readers consecutively evaluated each polyp candidate in the list and used the same confidence scale used by the unblinded radiologist to assign each candidate a confidence score of 1–5. Immediately after 2D scoring, the 3D subvolume was turned on, and readers repeated the scoring process by using both 2D and 3D viewing. The 2D scores could not be altered in the combined 2D and 3D mode. Although the entire data set was available, readers were not allowed to enter new polyp candidates. Results for each data set and reader were automatically stored as a text file. After a data set was evaluated, the next data set was automatically loaded for 2D viewing.
Statistical Analysis
Results for each reader were compared with the reference standard. We generated conventional receiver operating characteristic curves with LABMRMC software (Charles E. Metz, University of Chicago, Chicago, Ill), which uses the Dorfman-Berbaum-Metz algorithm (23). We used this software to test an analysis of variance model: By comparing the estimated area under the receiver operating characteristic curve (Az), we evaluated the effects of reader and reading mode, as well as the interaction of reader and reading mode in overall accuracy.
To test how features of each polyp candidate affected interpretation, we used repeated-measures analysis of variance and analysis of covariance to determine whether various continuous and categorical variables influenced readers' confidence scores. We performed two sets of analyses, thus evaluating the effect of each feature on both TP acceptance and FP rejection separately. For both sets of analyses, we tested for (a) the overall effect of each feature on reader accuracy across 2D and combined 2D and 3D interpretations and (b) the interaction between image feature and reading mode (ie, if significant differences existed in the feature effect between 2D and combined 2D and 3D interpretations).
In both of these analyses, we assumed that each polyp candidate was an independent observation. To test the independence assumption for FP observations, we first calculated the mean reader rating for each FP observation; thereafter, we calculated the standard deviation of these average ratings for each case. We then created 10 random permutations of the mean rating data and calculated standard deviations of these random data points by using the same pattern that was used with the original data. We then performed Mann-Whitney tests to compare the original within-case standard deviations with the randomly generated ones. Strong intracase correlations indicated the standard deviations from each case should be smaller than those from the randomly generated permutations. The independence of TP observations will be addressed in the discussion section.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
Effect of Reading Mode
For polyps of all sizes, overall diagnostic performance—as reflected by the Az values—increased for each reader when readers used combined 2D and 3D viewing compared with the performance when they used 2D viewing alone (Fig 3). This is different from sensitivity for polyp detection in a free-search study, as readers were restricted to decision making among preidentified polyp candidates. With 2D viewing, readers achieved Az values of 0.93–0.95. Even with this high degree of accuracy with 2D viewing, the addition of 3D viewing improved accuracy across all readers, as Az values increased to 0.96–0.97 (P < .001). When sensitivity was considered in individual readers, 3D viewing significantly improved sensitivity in readers 1 (P = .003), 3 (P = .021), and 4 (P = .016). Sensitivity did not significantly improve in reader 2 (P = .065); however, reader 2 had the highest sensitivity with 2D viewing.
|
Effect of Features
TP acceptance.—For TP polyp candidates, three features showed significant association with reader accuracy: preparation quality, maximum polyp width, and polyp height (Table). No other features were significantly associated with accuracy.
|
Maximum polyp width significantly affected reader confidence (P = .038) in that polyps with a larger diameter were associated with higher reader confidence that they were correctly classified as TP lesions. There was no significant interaction between reading mode and maximum polyp width.
Polyp height was associated with higher reader confidence in classification of TP lesions, as polyps of greater height were significantly associated with higher confidence levels (P = .019). For height, however, there was a significant interaction with reading mode (P = .025), as polyp height was associated with higher reader confidence levels with 2D viewing (P = .002) but not with combined 2D and 3D viewing (P = .132).
FP rejection.—For FP polyp candidates, six features were significantly associated with accuracy: colonic segment, attenuation, surface smoothness, distention, preparation, and true identity of polyp candidates (Table).
The colonic segment in which an FP candidate was located was significantly associated with accuracy for rejection of FP candidates (P = .007). To evaluate this on a segment-by-segment basis, pairwise comparisons for mean reader confidence were made between each of the eight segments (cecum, 1.42; ascending colon, 1.28; hepatic flexure, 1.18; transverse colon, 1.14; splenic flexure, 1.41; descending colon, 1.17; sigmoid colon, 1.33; rectum, 1.32). The only significant intersegment difference was between the cecum and the transverse colon (P = .007), in that readers rated FP candidates in the cecum as being more likely to be TP polyps than FP candidates in the transverse colon.
The CAD algorithm we used does not preclude detection of the ileocecal valve; thus, 39 of the 64 cecal FP candidates were actually the ileocecal valve. In a setting of nonsystematic colon viewing (as in this study), readers could mistake the ileocecal valve for a polyp if they were unaware of the overall colon location. Reanalysis of the mean reader confidence data, excluding the 39 ileocecal valve candidates (cecum-ileocecal valve, 1.40), showed that the difference between the cecum and transverse segments was still significant (P = .042).
Several other features, including polyp candidate attenuation, surface smoothness, local distention, and preparation quality, also affected FP polyp classification. Attenuation was strongly associated with reader performance (P < .001), but there was no significant interaction between attenuation and reading mode (P = .19). Surface smoothness (P < .001), local distention (P = .034), and preparation (P < .001) were also associated with a reader's ability to reject FP polyp candidates; however, surface smoothness, local distention, and preparation did not show a significant interaction with reading mode (P = .054, P = .054, and P = .24, respectively). However, readers tended to incorrectly classify FP candidates as polyps with higher confidence when polyp candidates were more heterogeneous and the surface appeared irregular instead of smooth. In addition, readers tended to incorrectly classify FP candidates as polyps with higher confidence when there was relatively poor distention and colon preparation.
The true identity of FP polyp candidates was strongly associated with a reader's ability to reject them (P < .001). FP candidates in the current experiment consisted of retained stool (31 of 652 [4.8%]; mean reader confidence, 2.55), lipoma (one of 652 [0.2%]; mean reader confidence, 1.5), thickened haustral fold (565 of 652 [86.6%]; mean reader confidence, 1.17), diverticulum (one of 652 [0.2%]; mean reader confidence, 1.00), imaging artifact (four of 652 [0.6%]; mean reader confidence, 1.41), ileocecal valve (39 of 652 [6.0%]; mean reader confidence, 1.44), and small bowel (11 of 652 [1.7%]; mean reader confidence, 1.18). Readers most frequently falsely classified focal retained stool as a polyp; readers also falsely classified the ileocecal valve as a polyp. In neither case did combined 2D and 3D viewing lead to a significant improvement in accuracy (stool, P = .090; ileocecal valve, P = .371). In contrast, readers more accurately classified thickened haustral folds as FP findings, and combining 3D and 2D viewing significantly improved reader performance in this task (P < .001).
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
Effect of Features
Acceptance of TP findings.—A significant association between the quality of colon preparation and the likelihood of a reader accurately accepting TP polyp candidates was confirmed. It has been written that a well-prepared colon is necessary for maximal diagnostic performance of CT colonography (18,24). We observed that reader confidence in acceptance of TP polyp candidates was significantly better when the colon preparation was best. Conversely, readers demonstrated decreased confidence scores when there was retained fluid or stool. Thus, colon preparation remains a critical determinant of the diagnostic performance of CT colonography.
We observed that the maximum diameter and the height of TP polyp candidates were associated with increased reader confidence. This result is not surprising, given the well-established pattern in which the diagnostic performance of CT colonography is directly related to the size threshold for lesion detection (5). In considering polyp height alone, however, we found that combined 2D and 3D reading significantly helped readers correctly classify short lesions. This is not a surprising result, since the accentuated shading and viewing available with 3D viewing would be expected to highlight these subtle lesions. Note that this effect was shown for small sessile lesions but not necessarily for large-diameter flat lesions (26), as the latter were not represented in our candidate pool.
Rejection of FP findings.—Even though we examined a polyp-enriched cohort, there were approximately 10 FP candidates for every TP candidate. This afforded greater statistical power to discern significant relationships between FP candidate features and reader confidence.
As with TP candidates, with FP candidates, preparation was strongly associated with accuracy in rejecting FP findings. In cases where local preparation was suboptimal, readers became less certain that a structure with a polypoid appearance could be discarded. This observation also held for distention, in that readers were less likely to discard an FP finding associated with poor distention. Two perhaps counterintuitive results were that readers were more likely to reject FP candidates if they were rated as smoother or if they were of more homogeneous attenuation. In the literature, it is well documented that retained stool and thickened haustral folds are the main sources of FP findings (27). To distinguish stool from a true polyp, heterogeneity in attenuation (particularly foci of entrapped gas) and an irregular or "geometric" surface are useful features. A limitation of our study was that readers used fixed window and level settings. Providing interactive window and level adjustment may have enabled better evaluation (21,28,29). Another reason we believe readers tended to correctly reject FP candidates with a smooth appearance is that the majority of such candidates were haustral folds, which have a smooth appearance and homogeneous attenuation.
Looking further at the FP candidates in terms of retained stool versus thickened haustral folds, the type of structure had a significant effect. In particular, the addition of 3D viewing significantly improved readers' accuracy in classifying an FP candidate in the case of thickened haustral folds; this improvement lends further support to this concept in the setting of preidentified polyp candidates. Regarding retained stool, we did not provide readers with prone and supine images together or use stool tagging and thus limited these potential aids in classification (5).
Study Limitations
Our study had several limitations. Use of only supine or prone data, lack of stool and fluid tagging, and inability to adjust window and level settings were previously mentioned. Also, the results obtained with the SURF platform may not be generalizable to other platforms. Experimental polyp candidates were generated by simulating an algorithm that would include all TP candidates within an arbitrary total of 15 candidate lesions. The composition of candidate lists could vary with other detection algorithms, and reader fatigue or bias could affect the results. This effect would depend on the total number of candidates or cases viewed.
Our study was also limited by the number of polyps, polyp size distribution, use of an overall polyp-enriched cohort, and total number of test cases. The limited number of TP polyps limited our power to discern relationships between candidate features and reader confidence.
Similarly, the statistical analyses we used did not take into account multiple candidates per data set. While our analysis of the intracase standard deviations for mean ratings of FP candidates lends support to the independence assumption for FP polyp candidates, there may have been more subtle correlations in the data that were not detected with this method. In addition, the small number of TP polyps precluded explicit testing of the independence assumption for TP polyps. Nearly three-quarters of patients with positive findings contributed only one or two positive lesions to the sample. Approximately half of patients with positive findings contributed only one lesion. From a biologic perspective, however, in a population like ours without familial adenomatous polyposis, each true polyp within a single patient is histologically distinct in that each arises from individual cells. While there may be genetic predispositions or environmental factors that contribute to similar origins of polyps in the same patient, biologic findings generally support the independence assumption for TP candidates. Nonetheless, we acknowledge that the use of statistical methods that do not take intracase clustering into account was a potential limitation of the study.
Other limitations include the fact that some of the features characterized in the current study are difficult to control systematically. For preparation, retained fluid and stool could be evaluated separately. Similarly, surface smoothness is a qualitative feature that might cover a range of characteristics. In addition, we were not able to compare 2D interpretation time with combined 2D and 3D interpretation time, as combined 2D and 3D interpretation was performed immediately after 2D interpretation, and case memory would likely affect interpretation time. It is not known if the addition of 3D viewing would increase or decrease classification time; however, answering this question would require a different experimental design.
An additional limitation is the fact that we examined this cohort with various tube currents (30) and that image noise at different tube current settings could affect lesion appearance, CAD performance, or both. Another limitation is that all readers had considerable (>100 cases) experience with CT colonography; thus, our results may not be generalizable to less-experienced radiologists.
In conclusion, we found that the addition of 3D viewing improved reader accuracy in the classification of preidentified polyp candidates and that a subset of candidate features, as well as local colonic characteristics, affected reader accuracy in polyp candidate classification. As preliminary data, these observations suggest that 3D viewing remains important for maximizing performance of human readers, even when CT colonography interpretation is augmented by CAD.
|
ADVANCES IN KNOWLEDGE |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
|
FOOTNOTES |
|---|
Abbreviations: Az = area under the receiver operating characteristic curve • CAD = computer-aided detection • FP = false-positive • SURF = Stanford University Radiology Framework • 3D = three-dimensional • TP = true-positive • 2D = two-dimensional
Authors stated no financial relationship to disclose.
Author contributions: Guarantors of integrity of entire study, R.S., S.N., C.F.B.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; literature research, R.S., C.F.B.; clinical studies, R.S., E.W.O., M.E.Z., D.M., D.S.P., C.F.B.; experimental studies, R.S., S.N., E.W.O., R.B.J., J.Y., D.S.P., A.J.S., C.F.B.; statistical analysis, P. Schraedley-Desmond, D.S.P.; and manuscript editing, all authors
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONADVANCES IN KNOWLEDGEReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  D. Hock, R. Ouhadi, R. Materne, A.-S. Aouchria, I. Mancini, T. Broussaud, P. Magotteaux, and A. Nchimi Virtual Dissection CT Colonography: Evaluation of Learning Curves and Reading Times with and without Computer-aided Detection Radiology, September 1, 2008; 248(3): 860 - 868. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. M. Summers, L. R. Handwerker, P. J. Pickhardt, R. L. Van Uitert, K. K. Deshpande, S. Yeshwant, J. Yao, and M. Franaszek Performance of a Previously Validated CT Colonography Computer-Aided Detection System in a New Patient Population Am. J. Roentgenol., July 1, 2008; 191(1): 168 - 174. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
N. Petrick, M. Haider, R. M. Summers, S. C. Yeshwant, L. Brown, E. M. Iuliano, A. Louie, J. R. Choi, and P. J. Pickhardt CT Colonography with Computer-aided Detection as a Second Reader: Observer Performance Study Radiology, December 1, 2007; 246(1): 148 - 156. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
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 | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
