Does a computer-aided detection algorithm in a second read paradigm enhance the performance of experienced computed tomography colonography readers in a population of increased risk?
- First Online:
- Cite this article as:
- de Vries, A.H., Jensch, S., Liedenbaum, M.H. et al. Eur Radiol (2009) 19: 941. doi:10.1007/s00330-008-1215-3
We prospectively determined whether computer-aided detection (CAD) could improve the performance characteristics of computed tomography colonography (CTC) in a population of increased risk for colorectal cancer. Therefore, we included 170 consecutive patients that underwent both CTC and colonoscopy. All findings ≥6 mm were evaluated at colonoscopy by segmental unblinding. We determined per-patient sensitivity and specificity for polyps ≥6 mm and ≥10 mm without and with computer-aided detection (CAD). The McNemar test was used for comparison the results without and with CAD. Unblinded colonoscopy detected 50 patients with lesions ≥6 mm and 25 patients with lesions ≥10 mm. Sensitivity of CTC without CAD for these size categories was 80% (40/50, 95% CI: 69–81%) and 64% (16/25, 95% CI: 45–83%), respectively. CTC with CAD detected one additional patient with a lesion ≥6 mm and two with a lesion ≥10 mm, resulting in a sensitivity of 82% (41/50, 95% CI: 71–93%) (p = 0.50) and 72% (18/25, 95% CI: 54–90%) (p = 1.0), respectively. Specificity without CAD for polyps ≥6 mm and ≥10 mm was 84% (101/120, 95% CI: 78–91%) and 94% (136/145, 95% CI: 90–98%), respectively. With CAD, the specificity remained (nearly) unchanged: 83% (99/120, 95% CI: 76–89%) and 94% (136/145, 95% CI: 90–98%), respectively. Thus, although CTC with CAD detected a few more patients than CTC without CAD, it had no statistically significant positive influence on CTC performance.
KeywordsColon CT colonography Computer-aided detection Virtual Colonoscopy Flat lesions
Computed tomography colonography (CTC) has consistently shown to have a high accuracy for colorectal neoplasia, and has recently been included in the official guidelines for colorectal cancer screening .
In the past years, efforts have been made to in order to increase its accuracy, e.g., labeling fecal material with a contrast agent (fecal tagging), automatic insufflation and improvement of workstations. Despite these efforts, visible lesions are still missed, even by well-trained radiologists.
Computer-aided detection (CAD) is a promising technique [2, 3, 4, 5] that could be helpful in reducing these false-negative findings [6, 7]. However, even if the CAD performance would be excellent, it does not automatically translate into equivalent reader performance [8, 9], i.e., CAD hits can be disregarded by the observer. This stresses the complex interaction between CAD and the observers.
Recent studies concluded that in a selected population CAD significantly improved per-polyp sensitivity for less experienced observers [10, 11, 12, 13]. Though experienced observers benefited proportionally less from CAD [14, 15]. Therefore, the potential increase by CAD in accuracy for experienced observers is still controversial.
The additional value of CAD was tested in a selected and polyp-enriched population only. This may have a positive effect on the observer performance since observers may be more easily triggered to detect polyps. Secondly, the a priori chance that a finding is indeed a polyp has increased. Therefore, the additional value of CAD (that will have a similar detection pattern irrespective of the population) may be larger.
To our knowledge, the effect of CAD on the performance of observers has not been prospectively evaluated in an unselected patient population of increased risk for colorectal cancer. Therefore, the purpose of this study was to determine whether CAD in a second read paradigm could improve the performance characteristics in a practical setting. Based on indirect comparison of two experienced observers and CAD , we hypothesized that CAD could still improve experienced observer performance.
Materials and methods
The institutional review board of both hospitals approved the study. All patients gave written informed consent.
Consecutive patients with a personal or family history of colorectal polyps or cancer were invited to participate from February 2006 until July 2007. All patients were scheduled to undergo a routine colonoscopy at one or other of the two participating hospitals. Exclusion criteria were: age under 18 years, pregnancy, personal history of inflammatory bowel disease, familial adenomatous polyposis, Peutz-Jeghers syndrome, hereditary non-polyposis colorectal cancer, prior allergic reaction to iodine contrast, untreated hyperthyroidism, known colorectal polyps that were not removed at an earlier endoscopy.
Patients ingested 4 l polyethylene glycol electrolyte solution (KleanPrep; Helsinn Birex Pharmaceuticals, Dublin, Ireland) on the day before and the day of the examinations. If contraindicated, other regimes were used. Patients ingested 50 ml oral iodine contrast (ioxithalamate, 300 mg ml−1) (Telebrix, Guerbet, Roissy, France) with each liter of polyethylene glycol electrolyte solution.
The table displays the baseline patient characteristics and procedural details of CTC (n = 170)
Age in years: mean ± SD
57 ± 12
Symptomatic (according to referring physician)
Changed bowel habits
History of colonoscopy or sigmoidoscopy 10 years prior to CTC
Without polyps ≥6 mm
With polyps ≥6 mm
With a polyp ≥10 mm
CT system used
20 mg butyl scopolamine bromidec
1 mg glucagond
Mean volume of insufflated CO2e
4.5 l (range: 2.2–7.6)
Mean scanner room examination time
21 min (range 12–35)
Number of complications
The CTC examinations were evaluated by one observer of a group of five observers; one board certified abdominal radiologist, two radiology residents (2nd and 4th year) and two radiology research fellows. The observers read the CTC examinations in a quiet environment not pressured to provide rapid reports, although they knew that the colonoscopy would be performed within 3 h. They were blinded to clinical data.
Table lists the number of read CTC cases per observer and observer experience with and without colonoscopic verification
No. of read cases
No. of endoscopic examinations with colonoscopic verfication
No. of endoscopic examinations without colonoscopic verficationb
300 CTC and 300 MRC
230 CTC and 30 MRCa
100 CTC and 240 MRC
Just prior to the study, all had passed a test of 25 selected CTC examinations  by scoring above a predefined per-polyp sensitivity threshold of 90%. In 12 of these 25 patients, 19 polyps ≥6 mm (one flat lesion) and 10 polyps larger than 10 mm could be detected.
CTC image analysis
The observers were blinded to the CAD results during the initial reading. All patients were evaluated with a primary three-dimensional (3D) method (Endo 3D Unfolded, ViewForum, Philips Medical Systems, Best, The Netherlands). This validated method  was used to increase surface visibility and reduce reading time.
Additional two-dimensional (2D) displays with instant on-screen correlation were used for problem solving. Stool subtraction software was not used. The observers digitally recorded size (mm), morphology (pedunculated, sessile, flat) and colon segment (cecum, ascending colon, transverse colon, descending colon, sigmoid colon or rectum).
If the observers identified CAD lesions that were not detected in the unassisted evaluation, these could be added to the initial list of findings.
Interpretation time and image quality
Interpretation times for the unassisted read and for the evaluation of CAD candidates were recorded with a stopwatch for both the prone and supine positions.
When the reading was completed, the observer assessed the degree of colonic distension and quality of the fecal tagging on a four-point Likert-scale (good, sufficient, moderate, poor). The overall quality of the examination was assessed as “diagnostic” or “non-diagnostic”. If the quality was assessed as “non-diagnostic” by the examining physician, the patient was excluded.
All patients underwent colonoscopy within 3 h after CTC. A gastroenterologist (>200 colonoscopies), or fellow or nurse under direct supervision of a staff member performed the colonoscopy with a standard colonoscope (CF-140L; Olympus, Tokyo, Japan). Chromoendoscopy or narrow-band imaging to improve flat polyp detection was not performed. Patients received on request midazolam (Dormicum, Roche, Basel, Switzerland) and fentanyl (Hameln Pharmaceuticals, Hameln, Germany) or propofol (Fresenius Kabi, Uppsala, Sweden) and fentanyl. The examination was digitally recorded.
Polyp characteristics (size, morphology and segmental location) were documented on a case record form by an attending research nurse. Polyp size was measured with open biopsy forceps (8 mm). The determination of the morphology of polyps was done by the gastroenterologist based on the endoscopic classification of superficial neoplastic lesions . In this classification, flat polyps were defined as lesions with a maximum height of 2.5 mm (closed cups of biopsy forceps). Segmental unblinding was performed for CT lesions 6 mm or larger. Histology was obtained at colonoscopy, except in those cases in which polyp removal was technically impossible or when material was lost during the procedure.
Determination of lesion status
Observers were instructed that only hyperplastic, adenomatous (advanced and not-advanced) and potentially malignant lesions were considered true-positive lesions. This qualification was based on the histology report or—if histology was not acquired—based on the endoscopic report.
For CTC, a polyp was considered true-positive, if: (1) its appearance resembled the corresponding polyp at colonoscopy, (2) its segment or adjacent segment corresponded with the reference standard segment and (3) the polyp size as estimated by the endoscopist corresponded with size as measured on CTC, considering a margin of error of 50%. Since the colonoscopy measurement is subject to inaccuracy [21, 22] this criterion could be overruled by the first two criteria.
Polyps ≥6 mm at colonoscopy that were not identified by the observer without or with CAD, were re-evaluated with knowledge of the colonoscopic findings by a research fellow with experience of more than 300 CTC examinations verified by colonoscopy. In this re-evaluation the nature of all detection errors ≥6 mm (false-negative findings) was assessed and differentiated between perception errors (visible in retrospect) and non-perception errors (lesions not visible in retrospect).
Lesions not confirmed by colonoscopy ≥6 mm (false positives) were assessed by consensus by two experienced research fellows (300 colonoscopy verified CTC). The consensus panel determined whether the finding was related to bowel preparation.
Based on a prior feasibility study , we expected a 15% increase in sensitivity. In order to determine a statistically significant increase of 15% for polyps ≥6 mm, at least 39 lesions were required. For this approach, a McNemar test with continuity correction and a p value of 0.05 to indicate statistical significance was used. Based on prior studies in this patient population, we assumed that the prevalence of patients with polyps ≥6 mm would be 25% . We therefore required a minimal number of 39/0.25 = 156 patients. The total number of patients determined was 170.
Outcome parameters per patient
Sensitivity, specificity, positive- and negative-predictive values of CTC without and with CAD were calculated. Sensitivity and number of false-positive findings were calculated for CAD without interaction of the observers (stand alone). Furthermore, sensitivity was calculated for unblinded colonoscopy. The outcome parameters were determined for polyps ≥6 mm and ≥10 mm.
A patient was considered true-positive if CTC detected at least one polyp seen at colonoscopy, based on the matching criteria described previously. A patient was categorized as false negative if CTC detected no polyps (although present at the reference standard) or only those of a lower size category in comparison to the reference standard.
We used the McNemar test to compare per-patient sensitivity and specificity values between CTC without and with CAD.
Outcome parameters per polyp
We calculated the per-polyp sensitivity for CTC without and with CAD, CAD (stand alone) and blinded colonoscopy for lesions 6–9 mm and ≥10 mm. In this study, more than one polyp was detected in some patients. Therefore, generalized estimating equations (GEE) (SPSS, 15.0, Statistics, Chicago, USA) was used to revise the data clustering and dependency. In the GEE, the adjusted confidence intervals with regard to per-polyp sensitivity for CTC and blinded colonoscopy (i.e., before unblinding of CTC results) were assessed for CTC without and with CAD. In this same GEE method, regression analyses were done to compare the sensitivity values.
Outcome time parameters
The median interpretation time of the CTC reading without CAD and the median time to evaluate all CAD results were calculated.
Prevalence of flat polyps stratified for endoscopic colon examination
Because a relatively high number of flat polyps were detected in this population we retrospectively determined whether a colon examination 10 years or less prior to the CTC in the patient’s history could effect the prevalence of these polyps. The rationale for this retrospective study was an article published by MacCarty et al.  that suggested a higher prevalence of polyps in patients who had undergone a previous endoscopic colon examination. We did not specify the type of colon examination in colonoscopy, sigmoidoscopy or proctoscopy because it was not always clear which part of the colon was examined. The arbitrary period of 10 years was chosen since we assumed this would be the period from a polyp to grow into a tumor and a colon examination executed earlier may have effect the prevalence of flat lesions at CTC.
The prevalence of polyps in the group that had undergone a colon examination and the group that had not were compared with the McNemar test and stratified for size.
The degree of bowel distention was assessed as “good” or “average” in 161 patients (95%), “moderate” in eight patients (5%) and “poor” in one patient (1%).
Fecal tagging was assessed as “good” or “average” in 144 patients (85%), “moderate” in 24 (14%) and “poor” in two patients (1%).
Table displays the histology and morphology of polyps at seen and removed during colonoscopy
6–9 mm (n = 58)
≥10 mm (n = 30)
Adenomatous lesions classified as “advanced”
Adenomatous lesions classified as “not advanced”
Results on a per-patient and per-poly basis (%) (95% CI in parentheses)
CTC without CAD
CTC with CAD
80% (69–91) 40/50
82% (71–93) 41/50
64% (45–83) 16/25
72% (54–90) 18/25
84% (78–91) 101/120
83% (76–89) 99/120
94% (90–98) 136/145
94% (90–98) 136/145
68% (56–80) 40/59
66% (54–78) 41/62
64% (45–83) 16/25
67% (48–84) 18/27
91% (86–96) 101/111
92% (86–97) 99/108
94% (90–98) 136/145
95% (92–98) 136/143
83% (72–90) 48/58
84% (73–92) 49/58
60% (43–75) 18/30
70% (51–84) 21/30
Total number of false-positive lesions
Two patients were erroneously classified as having a lesion ≥6 mm after accepting a CAD hit and no patients without a lesion ≥10 mm were wrongly added to the list.
CAD on a stand-alone basis detected 74% (37/50) and 64% (16/25) of the patients with lesions ≥6 mm and ≥10 mm. There was no statistically significant difference between the observers and CAD in the respective size categories (p = 0.375 and p = 1.0). CAD had a median number of nine hits per-patient (25–75% quartiles: 5–15). Blinded colonoscopy detected 96% (48/50) and 100% (25/25) of the patients in the respective size categories. As displayed in Table 4, both the positive- and negative-predictive values of the observers with and without CAD were nearly unchanged.
The three polypoid perception errors 6–9 mm (Fig. 4) were missed because they were situated on a fold (n = 1), were clearly smaller than 6 mm when measured on CT (n = 1) or could be defined as flat on CT (n = 1). The two polypoid perception errors ≥10 mm were missed because they were situated on a fold (n = 1) or because of unclear reasons (n = 1).
Of the 53 false positive lesions detected by the reader without CAD, 23 (43%) findings were according to consensus related to bowel preparation. None of the four false-positive findings suggested by CAD and incorporated in the final list by the observer were related to bowel preparation.
Although CAD did not significantly increase sensitivity, it did not significantly alter specificity either: three extra false-positive lesions 6–9 mm and one extra lesion ≥10 mm in 170 patients were added to the list of the observer (Table 4).
CAD on a stand-alone basis detected 72% (42/58) of the polyps 6–9 mm and 60% (18/30) of the polyps ≥10 mm. Blinded colonoscopy detected 95% (55/58) of the polyps 6–9 mm and 100% (30/30) of the polyps ≥10 mm.
The observers had an a median interpretation time of 16 min 00 s (25–75% quartiles: 11 min 35 s-23 min 6 s) to complete the examination and a median time of 1 min 26 s (25–75% quartiles: 28 s-2 min 46 s) to evaluate all CAD results after the initial reading.
Prevalence of flat polyps stratified for endoscopic colon examination
Sixty-three percent (107/170) of the patients had undergone an endoscopic colon examination prior to CTC, 31% (53/170) had not. For ten patients, the history could not be retrieved. Of the polyps 6–9 mm, in patients with a history of endoscopy 27% (12/44) were flat, in contrast to 8% (1/12) of the polyps in patients without a history of endoscopy (p = 0.259). Of the polyps ≥10 mm, in patients with a history of endoscopy 60% (9/15) were flat, in contrast to 33% (4/12) of polyps in patients without a history of endoscopy (p = 0.168). Thus, the prevalence of flat polyps in both size categories was higher in the group that had undergone colon examination, though statistical significance was not reached.
Although CAD in a second-read paradigm detected one additional patient with a lesion ≥6 mm and two patients with a lesion ≥10 mm, it did not significantly improve per-patient sensitivity in this increased risk patient population.
Several CTC studies in which the additional value of CAD was evaluated (after the interaction with the observer) have reported good results in terms of polyp detection [12, 25, 26, 27]. All concluded that the observers detected statistically more polyps with CAD.
In contrast to these studies, we did not find a significant additional value for CAD. The study design of these studies differs from our study in a number of aspects: patient selection, inclusion and exclusion criteria and reference standard. However, we think that the most important difference between the aforementioned studies and our study is the fact that the observers in our study were more experienced, i.e., more than 100 CTC cases verified by colonoscopy. Since there is good evidence that experience in CTC results in less false-negative findings [28, 29], it is logical that it is more difficult to substantially increase the sensitivity of the observer with CAD. In studies that report data about the additional value of CAD on experienced observers [30, 31], experienced observers benefited proportionately less from CAD when compared with the inexperienced readers. This finding is supported by the results of this study.
Although a CAD algorithm has the potential to decrease the number of perceptual errors by exposing the observer to candidate lesions, it cannot account for interpretative errors. In the above-mentioned papers, a significant increase of false positives have been reported. Though the specificity in this study was not significantly increased, there were only two patients erroneously classified as having a lesion. Both false-positive lesions measured 6–9 mm, none was larger than 10 mm.
Even though the sensitivity of the observers was low (i.e., 72% for lesions ≥10 mm), CAD was not able to increase their performance. In our opinion, the reported sensitivity requires looking for causes in the population itself. All nine polyps ≥10 mm missed by the observer with CAD had a flat morphology (Fig. 4). These flat polyps are an important cause of false-negative findings [32, 33]. In this population, 13 of the 30 lesions ≥10 mm were flat and therefore an important explanation of the moderate sensitivity, not only for the observer but for CAD as well.
The unexpectedly high number of flat lesions may be related to the history of patients; MacCarty et al.  reported in a prospective study of 75 consecutive patients that more than 50% of the false negatives missed by experienced readers were not even visible in retrospect in a population that had been screened by colonoscopy 5 years prior to CTC. Nearly all these polyps were flat. In this study population, the prevalence of flat lesions was higher (although not statistically significant) in the group of patients that had undergone a colonoscopy or sigmoidoscopy prior to CTC as well (Table 1).
We concur with MacCarty and coworkers that previous screening could aversively affect CTC sensitivity in two ways: first, it is likely that many easy-to-see polyps would be detected and removed at the initial screening, and fewer hard-to-see polyps would be detected and removed; second, endoscopic polypectomy may have been incomplete. Remnants of polyps are flatter than the original intact lesions, and would, therefore, be more difficult to detect. So we think that the selection of patients has an important influence on the test characteristics.
The type of bowel preparation, i.e., extensive or reduced, with or without oral contrast (iodine and/or barium) may influence the performance in terms of polyp detection and number of false-positive findings of CAD; polyps can be covered by fecal material or fecal remains may simulate polyps. In this study, we used PEG as an extensive bowel preparation for colonoscopy, combined with Telebrix that has a laxative effect as well. Although we have not evaluated the nature of all CAD candidates, the additional value of CAD did not seem to be impaired by this type of bowel preparation used in this study, since none of the false-positive lesions incorporated in the final list of the observers were prep-related and only two of 22 false-negative findings were covered by fecal material (although we think this is not the reason why they were missed).
This study has limitations; due to the small time-frame between the CTC and colonoscopy, the patients could be read by only one observer out of a group of five different observers. Each observer had a different level of experience. Therefore, although no statistically significant difference in sensitivity was measured between the observers, and none of the observers had a significant improvement in performance after CAD (data not shown), the best-performing observers could have leveled out the sensitivity of the least performing observers. Still, the situation as described in this paper is similar to the practical setting of many hospitals; each examination will not be read by five different radiologists but by only one of a pool of experts.
Secondly, the level of experience of the group of five readers was relatively high. It is likely that the additional value of CAD would be larger if the data were read by a relatively inexperienced reader group [35, 36]. Therefore, our conclusion may not apply to relatively untrained readers.
Thirdly, we evaluated CAD using a primary 3D reading paradigm. Although the discussion as to whether to read the data in 2D or 3D is still not settled, a (slight) superiority in terms of polyp detection with 3D is reported by some studies [37, 38]. Since the reported sensitivities of a primary 2D paradigm tend to be lower, CAD may have a larger additional value when used in a 2D reading protocol.
Fourthly, the relatively large number of flat polyps may limit the generalization of the results of this study. However, if we leave out all 26 flat polyps, and we only consider the remaining 62 polypoid lesions, we still cannot demonstrate a significant contribution of CAD to the sensitivity of the observer. This number was still more than the 39 polyps that were needed to demonstrate a 15% sensitivity difference according to our power analysis.
In conclusion, although CTC with CAD in a second read paradigm detected a few more lesions than CTC without CAD, CAD has no statistically significant positive influence on CTC performance in an increased-risk population when used by a relative experienced group of observers.
We thank Prof. M.J. van de Vijver, MD, PhD for assessing the histology of the polyps. We thank Mrs. A.C. Dobben, MSc, PhD for her statistical support. This work was supported by a grant from Philips Medical Systems (Best, The Netherlands).
This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.