Real-Time Characterization of Colonic Polyps: A Multicenter Prospective Study Evaluating the CAD-EYE System in Screening Colonoscopies.

INTRODUCTION
Colorectal cancer is the third most common cancer worldwide and the second leading cause of cancer-related mortality (1). Polypectomy during screening colonoscopy significantly reduces colorectal cancer incidence and mortality by interrupting carcinogenesis (2). Currently, all resected precancerous lesions undergo histological analysis to assess invasiveness, incurring significant ecological and financial costs (3,4).
Most of the resected lesions, however, are small polyps (≤10 mm) that carry a no risk (hyperplastic polyps) or low risk of malignant transformation (5). Therefore, the routine histological analysis of such lesions seems to provide limited clinical benefit and could potentially be replaced by accurate optical diagnosis. Clinical strategies based on optical diagnosis have been proposed, including the “diagnose-and-leave” approach, where nonneoplastic diminutive polyps in the rectosigmoid colon are left in situ, and the “resect-and-discard” strategy, where adenomatous colorectal lesions of ≤5 mm are resected without subsequent histological analysis.
The American Society of Gastrointestinal Endoscopy (ASGE) and the European Society of Gastrointestinal Endoscopy (ESGE) have defined performance benchmarks for incorporating these strategies into clinical practice. These are detailed in the Preservation and Incorporation of Valuable Endoscopic Innovations recommendations of the ASGE (6) and the consensus on Simple Optical Diagnosis Accuracy of the ESGE (7). Nevertheless, substantial variability in optical diagnosis accuracy among physicians and across healthcare centers, particularly in community settings, has led to frequent failure to meet the recommended safety thresholds (8,9).
Artificial intelligence (AI) systems have emerged as valuable tools in gastrointestinal endoscopy. Some have been developed for adenoma detection, demonstrating improved detection rates (10), while others have been tailored for the characterization of colorectal polyps, distinguishing “neoplastic” from “nonneoplastic” lesions to support clinicians in optical diagnosis.
Retrospective experimental studies have reported satisfactory diagnostic performance for AI systems in the histological characterization of lesions (10,11). To date, 9 prospective studies have evaluated the performance of AI-based histological characterization systems during nonmagnified colonoscopy, with inconsistent results. Some of these studies have specifically focused on diminutive rectosigmoid polyps (DRSPs).
The objective of this study was to evaluate the diagnostic performance of the CAD-EYE automatic characterization system for colorectal polyps in comparison with histological analysis.

MATERIALS AND METHODS

Study design
We conducted a cross-sectional, multicenter diagnostic study across five French centers (4 academic centers, 1 private center) with 19 endoscopists: the University Hospital of Bordeaux, University Hospital of Angers, Saint-Antoine Hospital, Georges Pompidou European Hospital, and American Hospital of Paris (see Supplementary Table 1, Supplementary Digital Content, http://links.lww.com/CTG/B452). The study was performed between October 2021 and January 2023. The protocol was approved by the Ethics Committee (CPP 1387 DM2) and conducted in accordance with the Declaration of Helsinki. The study was registered on ClinicalTrials.gov (NCT04921488). The study was funded by a grant from Bordeaux University hospital (“Appel d'offre interne du CHU de Bordeaux”). All patients provided informed written consent for participation.

Study population
Adult patients with an indication for screening colonoscopy were eligible to participate. This included individuals with a positive fecal immunochemical test, a personal or family history of colorectal cancer, or colorectal adenomas. The exclusion criteria were a guardianship or protection status, pregnancy, lack of health insurance, illiteracy, inability to undergo adequate colonic preparation, and a personal history of inflammatory bowel disease.
Patient screening was conducted during the precolonoscopy consultation. Only patients for whom at least 1 polyp was detected and resected during the colonoscopy, and then submitted for histological analysis, were included.

Colonoscopy procedures
Nonmagnified colonoscopies were performed using Fujifilm 760 endoscopes (Fujifilm, Tokyo, Japan). All procedures were the same in each center, conducted in accordance with the standard protocol. The Boston Bowel Preparation Scale (BBPS) was recorded for each colonic segment (left, transverse, right). Endoscopists were categorized according to their experience (≤5 or >5 years), evidence-based cutoff to distinguish early-career from experienced endoscopists (12).
When polyps were identified, complete excision was mandatory. The withdrawal time was calculated by subtracting the time spent on resections. All resected polyps were submitted for histopathological analysis. Hyperplastic polyps in the rectosigmoid region could be left in situ if the investigator was confident in the diagnosis, in line with European guidelines (7).

AI system
The CAD-EYE system (version 1; Fujifilm), a commercial real-time convolutional neural network-based AI tool, was used in this study. The detection component of the system highlights potential polyp regions by displaying blue squares on the screen and ringing a tone, identifying areas with visual characteristics consistent with polyps. False-positive detections may include folds, food residues, or stool remnants. The characterization component provides a binary output (“hyperplastic” or “neoplastic”) when used in blue light imaging mode.
This automatic system was activated at the start of colonoscopy withdrawal, displaying information on a separate screen with sound off to ensure the operator remained blinded to the CAD-EYE system's outputs. These outputs were recorded by a clinical research assistant (CRA) or an endoscopy nurse. A video recording of the colonoscopy withdrawal phase was created for each patient, both with and without the use of CAD-EYE, for research purposes (“proofreading”). These recordings were anonymized and stored securely on a cloud-based platform.

Polyp characterization and experimental procedure
Each detected polyp was visually assessed by the operator using white light and blue light imaging. If a polyp was identified only by the CAD-EYE system, the CRA or nurse was required to notify the operator within a reasonable timeframe. The morphology (based on the Paris classification), predicted size, location, predicted histology (adenoma with low-grade dysplasia, adenoma with high-grade dysplasia, cancer, hyperplastic lesion, sessile serrated lesion or SSL, or other), and method of resection were recorded on a standardized form, with each polyp assigned a unique identification number. After this, the CAD-EYE system was activated to provide a histological prediction (“hyperplastic” or “neoplastic”), which was immediately documented by the CRA (Figure 1).

Pathological analysis
Polyps were resected, placed in separate jars and sent for histological analysis. The pathologists were blinded to both the operator's and AI system's diagnoses. All polyps were assessed according to the Vienna classification. No centralized proofreading was performed. “Neoplastic” lesions included adenomas with low-grade or high-grade dysplasia, adenocarcinoma, and SSLs with dysplasia. “Nonneoplastic” lesions included hyperplastic polyps, SSLs without dysplasia, and “other.”

Proofreading of video recordings
A video recording of the colonoscopy withdrawal phase was created for each patient. Two gastroenterologists independently reviewed these recordings; one reviewed the CAD-EYE recordings, and the other reviewed the recordings without CAD-EYE. They had to determine whether the video was fully evaluable (i.e., whether all segments were visible), record the BBPS, document the number of polyps detected and resected in each segment, and for the reviewer of the CAD-EYE recordings, note the number of false positive detections identified by the AI system in each segment.

Study endpoints
The primary endpoint was the sensitivity of the CAD-EYE system for characterizing the neoplastic polyps, with histological analysis serving as the reference test. The secondary endpoints were the specificity, positive predictive value (PPV), and negative predictive value (NPV) of the CAD-EYE system for diagnosing the neoplastic nature of colorectal polyps, again using pathological analysis as the reference standard.
In addition, we compared the sensitivity of CAD-EYE and the investigator for detecting neoplastic features of colorectal polyps, using pathological analysis as the reference standard. Finally, we calculated the polyps per colonoscopy (PPC) rate, defined as the average number of polyps detected per patient, comparing colonoscopies assisted by CAD-EYE vs those performed without CAD-EYE. This analysis was conducted by 2 blinded expert endoscopists who reviewed the video recordings. Only videos capturing the entire colonoscopy procedure were included in this comparison.

Statistical analysis and sample size calculation
The sample size calculation was based on the sensitivity of the CAD-EYE system in characterizing neoplastic lesions. It was hypothesized that this sensitivity would be 96% (matching the sensitivity of the optical diagnosis made by the gastroenterologist during colonoscopy) and would exceed 85%. Assuming a one-sided error of 2.5%, a power of 80%, a 5% rate of missing data, and an estimated proportion of patients with at least 1 neoplastic polyp of 50%, the required sample size was calculated to be 140 patients. Given the expectation that 30% of patients would have no polyps identified during colonoscopy (and would therefore not be included), 200 patients were enrolled in the study. Statistical analysis was performed at the polyp level and based on available data.
To estimate the sensitivity of the CAD-EYE system in characterizing neoplastic lesions, neoplastic polyps identified by the investigator during colonoscopy and confirmed by histological analysis were analyzed. To estimate the specificity of the CAD-EYE system, nonneoplastic polyps identified by the investigator during colonoscopy and confirmed by histological analysis were analyzed. The PPV and NPV were calculated for polyps classified as neoplastic and benign, respectively, by the CAD-EYE system.
The 95% CIs were generated for each measure. Comparisons with the theoretical value of 85% were performed using the χ2 test or Fisher exact test if the assumptions for the χ2 test were not met. The number of polyps detected per patient using the CAD-EYE system was compared with that observed by the gastroenterologist during colonoscopy using a paired Student t test.

RESULTS

Baseline characteristics
During the study period, 181 patients were deemed eligible for enrolment across the five participating centers. Six patients were excluded: 3 because of incomplete colonoscopy and 3 for protocol deviations. Thirty-six patients were excluded because of absence of polyps or missing histology, leaving 139 patients for primary analysis (Figure 2).
Most of the participants were male (66%), and their median age was 66 years (Table 1). The main indication for colonoscopy was a personal history of colorectal adenoma (35%). Bowel preparation was adequate (BBPS ≥6) in 92% of cases. Of the 19 operators, 9 had less than 5 years of experience (see Supplementary Table 1, Supplementary Digital Content, http://links.lww.com/CTG/B452). In total, 398 polyps were detected, resected, and analyzed by pathologists. All baseline characteristics are detailed in Table 1.

Study outcomes

CAD-EYE diagnostic performance for polyp characterization.
In total, 358 polyps were included in the analysis because polyps classified as “undetermined” by pathologists (n = 40) were excluded (food residues, normal colonic mucosa, etc.). The blinded pathologists classified 245 (68%) polyps as neoplastic and 113 (32%) polyps as nonneoplastic. Optical diagnosis by endoscopists classified 252 (70%) polyps as neoplastic, 104 (29%) polyps as nonneoplastic, and 2 (1%) polyps as undetermined. CAD-EYE characterization was feasible in 96% of cases: 15 polyps classified as undetermined (4%), 210 as neoplastic (59%), and 133 as nonneoplastic (37%).
The sensitivity of the CAD-EYE system for characterizing the neoplastic nature of polyps did not significantly exceed the predetermined threshold of 85% (sensitivity, 0.80; 95% CI, 0.74–0.85; P = 0.064). All performance parameters of the CAD-EYE system for predicting neoplastic lesions are detailed in Table 2.
There was no variability in CAD-EYE sensitivity based on the location of polyps (Table 3). CAD-EYE sensitivity increased with larger polyp size (see Supplementary Tables 2 and 3, Supplementary Digital Content, http://links.lww.com/CTG/B452). Sensitivity remained consistent even when SSLs without dysplasia were classified as neoplastic; the sensitivity of the CAD-EYE system still did not significantly exceed the 85% threshold (sensitivity, 0.76; 95% CI, 0.70–0.81) (see Supplementary Table 4, Supplementary Digital Content, http://links.lww.com/CTG/B452).
The CAD-EYE specificity, PPV, and NPV tended to be higher for rectosigmoid polyps (Table 3). Specificity was greater for polyps < 10 mm than for those ≥ 10 mm, while PPV and NPV were higher for polyps ≥ 10 vs < 10 mm (see Supplementary Table 3, Supplementary Digital Content, http://links.lww.com/CTG/B452). These values were unaffected when SSLs without dysplasia were considered neoplastic (see Supplementary Table 4, Supplementary Digital Content, http://links.lww.com/CTG/B452).
The diagnostic performance of the CAD-EYE system was notably improved when focused on DRSPs only. The sensitivity, specificity, and NPV of the CAD-EYE system was, respectively, 0.80 (95% CI, 0.61, 0.92), 0.97 (95% CI, 0.87–0.99), and 0.86 (95% CI, 0.73–0.95) (Table 4).

Endoscopists' diagnostic performance for polyp characterization.
Endoscopists' sensitivity for predicting the neoplastic histology of colorectal polyps was significantly higher than that of the CAD-EYE system (P = 0.001). All diagnosis performance parameters are detailed in Table 2. Diagnosis performance tended to be superior for endoscopists with > 5 years of experience, except for sensitivity, which showed no significant difference (see Supplementary Table 5, Supplementary Digital Content, http://links.lww.com/CTG/B452).

Performance of polyp detection
During the 139 colonoscopies, 420 polyps were detected by endoscopists, including 227 adenomas. The adenoma per colonoscopy rate was 1.6. Of the 398 polyps that were ultimately resected and analyzed, 19 (5%) were initially detected by the nurse or CRA using CAD-EYE on the auxiliary screen. Their characteristics are detailed in Supplementary Table 6 (see Supplementary Digital Content, http://links.lww.com/CTG/B452).
We compared polyp detection during standard colonoscopies without CAD-EYE vs detection during video reviews, with and without CAD-EYE assistance. Only videos that captured the entire colonoscopy were included in this analysis (n = 48). The PPC rate was significantly higher for CAD-EYE-assisted colonoscopies than for standard colonoscopies (3.3 and 2.3, respectively; P < 0.001).
There was no significant difference in polyp detection based on physicians' experience. However, the PPC rate was higher when bowel preparation quality was better (BBPS of ≥ 6 vs < 6) according to video review without CAD-EYE assistance (3.6% vs 1.9%, P = 0.012) (see Supplementary Table 7, Supplementary Digital Content, http://links.lww.com/CTG/B452).

False detections
CAD-EYE yielded a mean of 5 false detections in both the right and transverse colon, and 5.5 in the left colon.

DISCUSSION
This cross-sectional, multicenter study evaluated the diagnostic performance of the CAD-EYE system for predicting the neoplastic histology of colorectal polyps. CAD-EYE characterization was feasible in 96% of cases. CAD-EYE sensitivity was 80% (95% CI, 0.74–0.85), falling short of the predefined 85% threshold. By contrast, endoscopists' sensitivity was 90%, significantly exceeding that of the CAD-EYE system. The specificity and NPV of the CAD-EYE system were 79% (95% CI, 0.71–0.87) and 64% (95% CI, 0.55–0.72), respectively.
In screening colonoscopy, sensitivity and NPV are the most clinically relevant metrics because low values may result in missed neoplastic lesions and increased risk of interval cancers (13). The sensitivity and NPV observed in this study were suboptimal, suggesting that CAD-EYE is not yet ready for routine clinical implementation.
Nevertheless, all diagnostic performance parameters of the CAD-EYE system tended to be higher when the analysis was restricted to rectosigmoid polyps rather than polyps in the right and transverse colon. Hassan et al (14) reported that the accuracy of CADx, similar to that of endoscopists, was greater for rectosigmoid polyps than for proximal polyps. This disparity may be attributed to a lower concordance between the macroscopic and microscopic appearance of proximal colonic polyps, leading to less reliable optical diagnosis.
For DRSPs, the sensitivity, specificity, and NPV of the CAD-EYE system were high but still fell short of the thresholds required for the “diagnose-and-leave” strategy (80%, 97%, and 86%, respectively). These findings align with those of other studies evaluating the diagnostic performance of the CAD-EYE system without optical magnification (15–18). Indeed, studies focusing exclusively on DRSPs reported sensitivities of >80% (86.4% and 81.2%) but <90% (15,17). By contrast, studies analyzing polyps of varying sizes revealed lower diagnostic performance (15,20). When focusing on diminutive colonic polyps, CAD-EYE sensitivity and specificity were insufficient to meet the thresholds set by the ESGE and ASGE for the “resect-and-discard” strategy (73% and 79%, respectively).
Other studies have evaluated different systems. Satisfactory performance parameters were found for systems focusing on DRSPs (14,16,18,19), but less favorable results when all types of polyps were studied (16,18,21).
Rondonotti et al (17) evaluated the interaction between CADx and human factors in the decision-making process. Although the benefits of AI assistance seem to be marginal for experts, AI may help nonexperts improve their optical diagnostic performance during routine clinical practice (17). As demonstrated in other studies, CADx systems show promise but may be more effective as a second reader rather than an independent reader.
This study demonstrates encouraging results regarding physicians' sensitivity and NPV, surpassing that of CAD-EYE (NPV 77% vs 64%, respectively), but remaining insufficient for confidently ruling out neoplastic histology. These positive findings should be interpreted cautiously because the study was primarily conducted in expert centers. Furthermore, the specificity, NPV, and PPV of experienced endoscopists (>5 years of experience) tended to be superior to those of endoscopists with less experience. Other studies have shown that quality thresholds have been achieved by endoscopists with varying levels of experience (22) but have not been consistently met in community centers (8). This highlights the critical importance of high-quality training in the optical diagnosis of colonic polyps, both during initial traineeships and through continuing medical education. Physician certification in optical diagnosis would ensure expertise and enable strategies deployment.
In our study, the PPC rate was higher in CAD-EYE-assisted colonoscopies than in standard colonoscopies (3.3% vs 2.3%, P < 0.001), consistent with findings from other studies (23–25). Detection rates were higher when colonic preparation quality was better (BBPS ≥ 6), reinforcing the importance of proper colonic preparation. The adenoma per colonoscopy rate was 1.6, aligning with rates reported in similar studies (0.5–1.5), likely attributable to the eligibility criteria of patient cohort (screening colonoscopies, at least 1 polyp detection).
CAD-EYE false recognitions could potentially undermine clinicians' trust in the CAD-EYE system.
This study has several strengths. First, it was a multicenter study, which enhances the external validity of the results. Second, colonoscopies were performed under routine conditions, avoiding protocol-induced risks. Third, polyps of all sizes were analyzed, allowing broader evaluation.
This study has limitations. First, we did not assess physician-CAD-EYE interaction, although the system is intended as a secondary reader to support clinical decision-making. Second, SSLs were classified as nonneoplastic, reflecting CAD-EYE training based on their hyperplastic-like appearance, despite malignant potential. Reclassifying them as neoplastic lowered sensitivity to 76%. Focused analyses on SSLs may clarify system performance. Third, we performed a per-polyp analysis rather than a per-patient analysis, which limited our ability to draw conclusions about overall patient care. Finally, investigators could leave DRSPs in situ when confident of a hyperplastic diagnosis, reducing histological analyses and potentially lowering CAD-EYE performance metrics.
In conclusion, our study demonstrated that although the sensitivity of CAD-EYE system shows potential for predicting the histology of DRSPs, the required thresholds are still to be met. Endoscopists exhibited higher sensitivity than the CAD-EYE system for predicting neoplastic polyps. However, AI systems such as CAD-EYE seem to be beneficial for improving polyp detection during colonoscopies.

CONFLICTS OF INTEREST
Guarantor of the article: Arthur Berger, MD, PhD.
Specific author contributions: F.Z., A.B.: conceptualization. A.B.: data curation. X.D., A.G., E.C.-M., A.O., G.R., G.P., E.S., F.Z., A.B.: investigation. V.R., F.Z., A.B., A.B.: methodology. F.Z., A.B.: validation. M.M., A.B.: roles/writing - original draft - review & editing. Each author has approved the submitted version.
Financial support: Study funded by an internal call for proposals from Bordeaux University Hospital.
Potential competing interests: A.B.: Fujifilm, Olympus, Norgine. M.S.: Sandoz. X.D.: Co-founder and shareholder of Augmented Endoscopy; Consultancy for Norgine and Provepharma; Lectures for Abbvie, Alfasigma, Janssen, Mayoly, Medtronic, Norgine, Sandoz. E.C-M: Fujifilm, Cook, Abbvie. E.S.: Fujifilm, Boston Scientific. G.R.: consultant for Medtronic and Fujifilm.
Study Highlights

WHAT IS KNOWN
✓ There is substantial variability in colonic polyps optical diagnosis accuracy among physicians and centers. Artificial intelligence could help to standardize pratices.

WHAT IS KNEW
✓ In this study, CAD-EYE system performance parameters fell short of the threshold required for “diagnose-and-leave” and “resect-and-discard” strategies. Performance parameters were higher when analysis was restricted to diminutive rectosigmoid polyps.

✓ CAD-EYE system can improve polyp detection.

Supplementary Material