Comparison of a unidirectional panoramic 3D endoluminal interpretation technique to traditional 2D and bidirectional 3D interpretation techniques at CT colonography: preliminary observations

materials and methods Sixty-nine patients underwent computed tomography colonography (CTC) after bowel cleansing. Forty-five had no polyps and 24 had at least one polyp ≥6 mm. Patients underwent same-day colonoscopy with segmental unblinding. Three experienced abdominal radiologists evaluated the data using one of three primary interpretation techniques: (1) 2D; (2) bidirectional 3D; (3) panoramic 3D. Mixed model analysis of variance and logistic regression for correlated data were used to compare techniques with respect to time and sensitivity and specificity. Results Mean evaluation times were 8.6, 14.6, and 12.1 min, for 2D, 3D, and panoramic, respectively. 2D was faster than either 3D technique ( p < 0.0001), and the panoramic technique was faster than bidirectional 3D ( p = 0.0139). The overall sensitivity of each technique per polyp and per patient was 68.4 and 76.7% for 2D, 78.9 and 93.3% for 3D; and 78.9 and 86.7% for panoramic 3D. Conclusion 2D interpretation was the fastest overall, the panoramic technique was significantly faster than the bidirectional with similar sensitivity and specificity. The sensitivity for a single reader was significantly lower using the 2D technique. Each reader should select the technique with which they are most successful. Introduction Recent studies have demonstrated that computed tomography colonography (CTC) approaches optical colonoscopy in the detection of colorectal polyps ≥1 cm. 1–3 As a result of continued advances in the field of CTC, the American Cancer Society has now endorsed CTC as an approved method for colorectal cancer screening. 4 Given these recent results and endorsements, it is likely that the number of CTC studies performed will continue to increase. While maintaining a high sensitivity and specificity for polyp detection, reducing the evaluation time per study will be an important objective in order to meet the demand of the increased volume of studies. There continues to be some controversy regarding the optimal interpretation strategy with regard to lesion detection and time efficiency. A primary two-dimensional (2D) approach relies on scrolling through axial images and “tracking” the colon from rectum to caecum. If a lesion is seen using the 2D view, three-dimensional (3D) endoluminal imaging, multiplanar reconstruction (MPR) images, and adjustment of window/level settings are used for problem-solving. 5,6 Conversely, a primary 3D approach relies on performing initial fly-through navigations in the colon and using 2D images for problem-solving if an abnormality is detected during the fly-through. 7,8 Using a primary 3D approach requires a bidirectional navigation (antegrade and retrograde) in order to visualize potential lesions hidden by haustral folds. This increases data interpretation times when compared with a primary 2D technique. 1 However, a number of new 3D visualization techniques have been developed that could decrease the evaluation time by enabling more complete visualization of the colonic surface during a single fly-through or using virtual dissection techniques. 9,10 A commercially available “panoramic 3D technique” has been recently developed, which “opens up” colonic folds and reveals the colonic mucosa behind each fold, thereby theoretically allowing the reader to use a unidirectional approach. Using this technique, haustral folds are unfolded as the reviewer navigates through the colon thus revealing hidden areas of the colon. This could ultimately reduce total evaluation time when performing a 3D navigation as the primary interpretation technique. However, some distortion of the mucosa occurs during this navigation and the effect upon polyp detection and interpretation time is unclear. Therefore, the purpose of this study is to compare primary 2D, bidirectional 3D, and unidirectional panoramic 3D navigations of CTC studies with respect to interpretation time, sensitivity, and specificity. Methods Patients Institutional review board approval was obtained for this study, and it was HIPAA compliant. Signed informed consent was obtained from all participants. Sixty-nine patients were selected for inclusion in this retrospective study from a larger cohort of research participants undergoing same-day CTC and colonoscopy. Patients were included if they underwent same-day colonoscopy with segmental unblinding of the endoscopist to the CTC interpretation. This served as a reference standard for the presence or absence of colorectal polyps. The participants comprised 59 men and 10 women with a mean age of 62 years (range 50–82). Forty-five cases were negative (no polyps) and 24 were positive (showed polyps). In these 24 patients, a total of 30 polyps were present (10 were 6–9 mm and 20 were ≥10 mm). The indications for CTC in the 69 subjects was screening in 35 and for bowel-related symptoms in 34 (positive faecal occult blood test, change in bowel habits, or bleeding). All patients signed an IRB approved consent form. Patient preparation and CTC technique On the day prior to the study, patients ingested 4 l of a polyethylene glycol solution for bowel cleansing. Patients were asked to evacuate residual fluid/faecal material from the rectum just prior to the CTC examination. Neither faecal nor fluid tagging was used in this study. A rubber catheter was inserted into the rectum, and the colon was distended with carbon dioxide with the patient in the left lateral decubitus position and then slowly turned supine. Two litres of carbon dioxide were mechanically insufflated and a scout topogram was obtained to evaluate distension. If the topogram showed incomplete distension additional carbon dioxide was insufflated and the carbon dioxide continued to be insufflated throughout the procedure to maintain a pressure of 25 mmHg. Once complete distension was achieved (determined by visual assessment of the scout image), both supine (initially) and then prone CTC datasets were obtained using either a Siemens Sensation 64 ( n = 60) or 16-row CT machine ( n = 9; Siemens Medical Systems, Forcheim Germany) with 0.6–1.2 mm detector or 0.75–1.5 configuration, 1–1.5 mm section reconstruction, and overlapped every 1 mm. Additional CT parameters included 120 kVp and ≤50 effective mAs. The CT data were sent to a Siemens Leonardo workstation equipped with CTC software. Colonoscopy technique and reference standard Colonoscopy was performed immediately after CTC by a board certified gastroenterologist with 12 years experience in optical colonoscopy. In all cases, segmental unblinding of the endoscopist to the initial CTC interpretation was used as the reference standard. Segmental unblinding of the gastroenterologist was performed by having the CTC datasets immediately interpreted and the results given to a nurse coordinator who accompanied the patient to the endoscopy suite. All CTC data were prospectively interpreted by an abdominal radiologist (M.M.) with >10 years experience in CTC data interpretation. At endoscopy, upon withdrawal of the endoscope from the caecum, successive segments were interrogated by the endoscopist and correlated with the CTC findings. Two of the authors (M.M., D.K.L.) who were aware of the study design confirmed that all polyps detected by the reference standard could be seen on at least one of the two (supine or prone) CTC datasets. Data interpretation techniques The Siemens Leonardo workstation is equipped with CTC software that allows either a primary 2D, traditional bidirectional 3D (with 110° field of view) or a panoramic 3D interpretation technique. The panoramic view opens folds as they are passed, allowing the entire colonic mucosa to be visualized with a single endoluminal navigation. Specifically, the panoramic view combines five views (forward, up, down, left, and right) into one single circular display, while minimizing distortions at the seams. As each of these views has a field of view of 90°, the overall coverage is 270°. 11 The data were independently analysed by three abdominal radiologists who were unaware of the number of patients who harboured polyps. Two radiologists had experience in interpretation of at least 75 prior CTC datasets. These radiologists also completed a course where they had the opportunity to review 50 different CTC datasets with colonoscopy correlation with an experienced reader. The third radiologist had interpreted >100 CTC datasets prior to the initiation of this study. All three radiologists had been interpreting CTC examinations for at least 2 years. All three readers were experienced with traditional 2D and 3D interpretation techniques. None of the readers had any formal training using the panoramic 3D technique. Each radiologist reviewed a specific dataset using a pre-assigned method (2D, bidirectional 3D, or panoramic 3D) according to a randomized fractional factorial design. In order to satisfy constraints imposed by the statistical design and to mitigate confounding of between-reader differences with comparisons among interpretation techniques, four of the 45 negative (no polyps) cases and six of the 24 positive (cases with polyps) cases were read twice by each reader using two different techniques, for a total of 79 evaluations per reader. This allowed techniques to be compared using results from a single reader, thereby guaranteeing that differences among techniques could be estimated independent of any differences between readers. Furthermore, the design included instances where the same case was evaluated by two different readers using the same technique. As a result, it was possible to assess inter-reader variability. Overall, each reader interpreted between 25 and 28 cases using each of the three techniques. When interpreting an examination, all problem-solving techniques (2D, 3D, window/level adjustment) were available to the radiologist if an abnormality was detected. Polyp size and location and interpretation times were recorded for each evaluation. The findings at 2D, bidirectional 3D and panoramic 3D CTC interpretation were compared with those of the reference standard and with each other to determine performance characteristics of each interpretation technique. Polyp matching The reference standard size and location of all polyps was determined in consensus by the colonoscopist and one of the authors (M.M.) who was not participating in data interpretation for this study. This was based on review of endoscopic images and CTC data of all polyps. A polyp identified using one of the various CTC interpretation techniques was considered to be a true positive if the location was in the same or in the adjacent colonic segment as the colonoscopy and the size was within 3 mm on of the reference standard. If a polyp identified at CTC was not detected based on the reference standard, the polyp was considered a false positive. Statistical analysis Interpretation times The interpretation times were recorded by each reader using each technique. The interpretation time was the time in minutes to completely evaluate the colon using an individual technique and did not include the time to load the data into the colon application. The mean ± standard deviation of the evaluation time reported by each reader when using each technique was calculated. Mixed model analysis of variance (ANOVA) was used to compare the three CTC interpretation techniques with respect to evaluation time while accounting for systematic differences among readers and for the lack of statistical independence among interpretation times reported for the same patient. Sensitivity and specificit y The sample size was determined on the basis of a simulation analysis so that the study had 80% power to detect a 15 percentage point difference in per-patient sensitivity and a 10-point difference in per-patient specificity, so that the design satisfied the constraints underlying the fractional factorial design. More specifically, the sample size was determined by simulating the data for a given reader as random draws from a multivariate distribution, with each draw representing simulated diagnoses as positive versus negative for a given patient using each of the three techniques. In particular, when computing power for per-patient specificity (or sensitivity), the underlying distribution that was sampled specified that each 3D technique had a specificity of 95% (or 90%), while the 2D technique had a specificity of 85% (or 75%), that the 2D technique would agree with each 3D technique for 85% (or 80%) of patients and that the 3D techniques would be concordant 95% (or 90%) of the time. For each sample size, 1000 samples were simulated and then each sample was analysed using logistic regression for correlated data as described in the statistical analysis section. The results together with the assumption that approximately two-thirds of patients would be have negative results and the constraints of the fractional factorial design led to a sample size of 69 patients. Logistic regression for correlated data was used to compare interpretation techniques with respect to sensitivity and specificity. Specifically, generalized estimating equations based on a binary logistic regression model were used to compare the techniques while accounting for differences among readers and the correlation among assessments provided for the same patient. Sensitivity was assessed on both a per-polyp and per-patient basis, whereas specificity was evaluated per-patient only. Sensitivity (per-polyp) of each technique was stratified for each reader by polyp size (polyps 6–9 mm and polyps ≥10 mm). For per-patient sensitivity and specificity, a subject was considered test-positive for polyps according to a given reader when using a given technique if at least one polyp was detected; the patient was considered test-negative otherwise. Per-patient sensitivity was calculated as the percentage of test-positive patients among the patients seen at reference standard to have at least one polyp. Per-patient specificity was calculated as the percentage of test-negative patients among the patients seen at reference to be free of polyps. All reported p -values are two-sided and were considered statistically significant when less than 0.05. SAS version 9.0 (SAS Institute, Cary, NC, USA) was used for all statistical computations. Inter-reader variability Using data from cases evaluated by two different readers using the same technique, kappa coefficients and the percentage of concordant evaluations were computed as measures of inter-reader agreement. Kappa coefficients and the percentage of concordant opinions was computed using data from cases evaluated by a single reader using two techniques to assess agreement between techniques. Agreement was interpreted as fair-moderate when kappa was less than 0.6, as substantial when kappa was between 0.6 and 0.8, and as nearly perfect when kappa exceeded 0.8. 12 Results Reference standard According to the reference standard, there were a total of 30 polyps in 24 patients (10 polyps were 6–9 mm and 20 were ≥10 mm). The remaining 45 patients had no polyps according to the reference standard. As described above, four of the 45 negative (no polyps) cases and six of the 24 positive (cases with polyps) cases were read twice by each reader using two different techniques, for a total of 79 evaluations per reader. Overall, each reader interpreted between 25 and 28 cases using each of the three techniques. The cases were selected such that for each interpretation technique, 10 cases showed polyps and between 15 and 18 cases were normal. Interpretation time The mean evaluation times ± standard deviations amongst all three readers for the primary 2D, bidirectional 3D and panoramic 3D techniques were 8.6 ± 4.1, 14.6 ± 5.5 and 12.1 ± 8.3 min, respectively ( Table 1 ). Interpretation time for primary 2D review was significantly lower ( p ≤ 0.0001) than either of the 3D techniques. The panoramic 3D technique was significantly faster ( p = 0.0089) than the bidirectional 3D technique by 2.5 min on average. Sensitivity Averaged over all three readers, the sensitivities of detecting any polyp ≥6 mm in a positive patient (per-patient sensitivity) were 76.7, 93.3 and 86.7% for primary 2D, bidirectional 3D, and panoramic 3D evaluations, respectively ( Fig. 1 ). The per-patient sensitivity of primary bidirectional 3D review was significantly higher than that of primary 2D review ( p = 0.038; Table 2 ). The sensitivities of detecting all polyps ≥6 mm based on the reference standard (per-polyp sensitivity) were 68.4, 78.9, and 78.9% for primary 2D, bidirectional 3D and panoramic 3D evaluations, respectively. There was no significant difference in per-polyp sensitivity among the three techniques ( p = 0.294 for 2D versus each 3D, p = 1.0 when comparing 3D techniques; Table 2 ). When considering each reader independently, the per-patient sensitivity of reader 1 for each of the three techniques was 40, 90, and 90% for primary 2D, bidirectional 3D and panoramic 3D evaluations, respectively ( Table 3 ). Reader 1 was significantly more sensitive on a per-patient basis using the bidirectional 3D ( p = 0.0266) and panoramic 3D techniques ( p = 0.0266) over the 2D technique. Readers 2 and 3 did not have any significant difference in per-patient sensitivity among the three techniques. On a per-polyp basis, the sensitivity of reader 1 for detecting all polyps ≥6 mm was 38.5, 75, and 76.9% for primary 2D, bidirectional 3D and panoramic 3D evaluations, respectively ( Table 4 ). Reader 1 was significantly more sensitive on a per-polyp basis using the 3D panoramic technique over the 2D technique ( p = 0.339), and there was a trend toward significance for increased sensitivity with bidirectional 3D as compared with the 2D technique ( p = 0.0745). Readers 2 and 3 did not show any significant difference in per-polyp sensitivity among the three techniques ( Tables 4 and 5 ). Specificity The per-patient specificity of each technique was 98, 89.8, and 98% for primary 2D, bidirectional 3D and panoramic 3D evaluations, respectively ( Table 6 ). There was no significant difference in per-patient specificity among the three techniques overall or stratified by reader. Inter-reader variability Measures of inter-reader agreement and agreement between techniques are summarized in Table 7 . The results suggest that agreement was only fair-moderate for two of the readers, but was substantial or nearly perfect for the other two pairs of readers. Agreement between techniques was substantial or nearly perfect for the 12 cases evaluated by a single reader using different techniques. Discussion Prior to 2003, most CTC data were interpreted using a primary axial 2D technique. A 2D approach has been shown to be sensitive and time efficient. 5,6,13 A review of CT colonography experts in 2004 showed that the vast majority (20/25, 80%) felt that the optimal method of data interpretation was a primary 2D approach. 14 However, two large studies demonstrated a poor sensitivity for large (>10 mm) polyp detection using a primary 2D approach. 15,16 Although there are several reasons why the sensitivity may have been decreased in these studies, reader training and experience were likely important contributing factors. More recently, there has been a shift towards primary 3D interpretations. 2,3,7 This is due, in part, to advances in 3D software development. Recent studies have demonstrated comparable sensitivity using a primary 3D approach at CTC to optical colonoscopy. 2,3 Proponents of a primary 3D data interpretation technique argue that it is more realistic, simulates conventional colonoscopy, and is more sensitive than a primary 2D technique. A 2007 study found significantly increased per-polyp sensitivity using a primary 3D approach when compared with a primary 2D approach. 8 A potential bias in this study was that the readers were trained in and used a primary 3D approach when reviewing their CTC data. Data interpretation times may be substantially increased when using a primary bidirectional 3D technique because polyps that are hidden behind haustral folds may be missed on a single navigation, necessitating two fly-throughs in the antegrade and retrograde direction. Moreover, the 2008 ACRIN trial showed that a primary 2D review is faster than a primary bidirectional 3D review, with overall similar sensitivity and specificity for adenomas ≥10 mm. 1 With the development of a panoramic 3D technique, the interpretation time should be reduced as, theoretically, only a unidirectional approach is necessary. The aim of the present study was to determine the performance characteristics of a primary 2D versus bidirectional 3D versus panoramic 3D interpretation algorithm among three different readers with expertise using both primary 2D and 3D interpretation techniques. The results of the present study indicate that a primary 2D technique is significantly faster than either 3D approach, with an average interpretation time of 8.6 min. The panoramic 3D technique (12.1 min), however, was significantly faster than the bidirectional 3D technique (14.6 min) by 2.5 min. Regarding polyps ≥6 mm, no significant difference was found in the per-polyp sensitivity of all three techniques. However, when considering per-patient sensitivity, the bidirectional 3D technique (93.3%) was significantly more sensitive than the 2D technique (76.7%). The panoramic 3D technique was also fairly sensitive (86.7%), but without statistical difference from the 2D technique. When each reader's per-patient and per-polyp sensitivity was evaluated separately, reader 1 was found to be significantly more sensitive on a per-patient basis using either of the 3D techniques over the 2D technique and was significantly more sensitive using the 3D panoramic over the 2D technique on a per-polyp basis. However, for readers 2 and 3 there was no significant difference between the three techniques with respect to sensitivity. The specificity of all three techniques was found to be quite high, ranging from 89.8% for the bidirectional 3D technique to 98% for the 2D and panoramic 3D techniques. These data indicate that for some readers (readers 2 and 3), a 2D approach may be more practical as the data can be interpreted faster and with similar sensitivity and specificity to a 3D technique. For other readers (reader 1), a 3D approach would seem mandatory as this reader had an overall poor sensitivity using a primary 2D approach. For a reader like this, the choice to use a panoramic 3D or bidirectional traditional 3D endoluminal approach can be predicated on user preference realizing that the panoramic technique can be performed with similar sensitivity and specificity but in less time. There are some limitations to this study. First, the sample size of 79 interpretations and 30 polyps for each reader was relatively small. However, we were able to show overall among the three readers statistically significant differences in interpretation times and in sensitivity between a primary 2D and 3D interpretation technique for one of the readers. Second, the possibility of recall bias exists in the 10 cases that were evaluated twice by each reader. However, there were no patient identifiers present, the cases were randomly evaluated, at least 2 weeks elapsed between interpretations, and primary 2D and 3D interpretation techniques are sufficiently different in their imaging appearance to make recall bias a minor factor. Third, neither faecal nor fluid tagging was used in this study. There is a trend towards more universal use of these agents in CTC as tagging may theoretically improve sensitivity and specificity. Currently, tagging is not part of the American College of Radiology (ACR) standards for patient preparation. It should be noted that despite the use of tagging in the ACRIN trial, the specificity was quite low in that study and lower than that observed in the present study without the use of tagging. 1 It is unclear how the use of tagging agents may have impacted on results of the present study. Finally, readers in the present study had only minimal prior experience with panoramic 3D evaluation. With further experience utilizing the panoramic 3D technique, the interpretation time would be expected to decrease. In addition, in future virtual dissection views and computer-aided detection will likely be an integral part of data interpretation. Nevertheless, current data interpretation techniques rely on either a primary 2D or 3D interpretation technique. In conclusion, the preliminary observations of the present study suggest a primary 2D interpretation technique to be significantly faster than either of the 3D techniques. Although the 3D panoramic technique is significantly more sensitive on a per-patient basis than the 2D technique, the other measures of sensitivity and specificity yielded no significant differences. In fact, the overall results were likely influenced by one of the three readers who had a significantly higher sensitivity using either of the two 3D techniques over the 2D technique, while the other two readers had similar sensitivities using all three techniques. Sensitivity, therefore, appeared to be quite reader-dependent. The results of the present study suggest that there may not be a universally optimal CTC interpretation technique, but rather readers should utilize the technique with which they are personally most successful. Determination of an individual's optimal interpretation technique could be determined at national CTC courses, teaching files that are being developed, or personal review of CTC cases with endoscopic confirmation. References 1 C.D. Johnson M.H. Chen A.Y. Toledano Accuracy of CT colonography for detection of large adenoma and cancers N Engl J Med 359 2008 1207 1217 2 D.H. Kim P.J. Pickhardt A.J. Taylor CT colonography versus colonoscopy for the detection of advanced neoplasia N Engl J Med 357 2007 1403 1412 3 S. An K.H. Lee Y.H. Kim Screening CT colonography in an asymptomatic average-risk Asian population: a 2-year experience in a single institution AJR Am J Roentgenol 191 2008 W100 W106 4 B. Levin D.A. Lieberman B. McFarland Screening and surveillance for the early detection of colorectal cancer and adenomatous polyps, 2008: a joint guideline from the American Cancer Society, the multi-Society task force on colorectal cancer, and the American College of Radiology CA Cancer J Clin 58 2008 1 31 5 A.H. Dachman J.K. Kuniyoshi C.M. Boyle CT colonography with three- dimensional problem solving for detection of colonic polyps AJR Am J Roentgenol 171 1998 989 995 6 M. Macari A. Milano M. Lavelle Comparison of time-efficient CT colonography with two-dimensional and three dimensional colonic evaluation for detecting colorectal polyps AJR Am J Roentgenol 174 2000 1543 1549 7 P.J. Pickhardt J.R. Choi I. Hwang Computed tomographic virtual colonoscopy to screen for colorectal neoplasia in asymptomatic adults N Engl J Med 349 2003 2191 2200 8 P.J. Pickhardt A.D. Lee A.J. Taylor Primary 2D versus primary 3D polyp detection at screening CT colonography AJR Am J Roentgenol 189 2007 1451 1456 9 K.T. Johnson C.D. Johnson J.G. Fletcher CT colonography using 3600 virtual dissection: a feasibility AJR Am J Roentgenol 186 2006 90 95 10 H. Hoppe C. Quattropani A. Spreng Virtual colon dissection with CT colonography compared with axial interpretaion and conventional colonoscopy: preliminary results AJR Am J Roentgenol 182 2004 1151 1158 11 B. Geiger C. Chefd'Hotel S. Sudarsky Panoramic views for virtual endoscopy MICCAI 2005 662 669 12 J.R. Landis G.G. Koch The measurement of observer agreement for categorical data Biometrics 33 1977 159 174 13 M. Macari E.J. Bini X. Xue Prospective comparison of thin-section low-dose multislice CT colonography to conventional colonoscopy in detecting colorectal polyps and cancers Radiology 224 2002 383 392 14 M.A. Barish J.A. Soto J.T. Ferrucci Consensus on current clinical practice of virtual colonscopy AJR Am J Roentgenol 184 2005 786 792 15 P.B. Cotton V.L. Durkalski B.C. Pineau Computed tomographic colonography (virtual colonoscopy)—-a multicenter comparison with standard colonoscopy for detection of colorectal neoplasia JAMA 291 2004 1713 1719 16 D.C. Rockey E. Paulson D. Niedzwiecki Analysis of air contrast barium enema, computed tomographic colonography, and colonoscopy: prospective comparison Lancet 365 2005 305 311