Society: AGA

Introduction: Although colonoscopy is considered the gold standard for detection and removal of premalignant polyps, up to 26% of lesions are missed in tandem studies. Computer-aided polyp detection (CADe) has shown promise in increasing polyp detection rates. The aim of this study was to evaluate a novel CADe system, ‘Magentiq Eye Automatic Polyp Detection System’ (ME-APDS), in a non-iFOBT screening and surveillance colonoscopy population.

Methods: A multicenter, randomized, controlled (RCT) trial was conducted at 10 hospitals in Europe, US and Israel. Patients (18-90 years), referred for screening (non-iFOBT) or surveillance colonoscopy, were included. Patients were randomized (1:1) to undergo CADe-assisted colonoscopy or conventional colonoscopy (CC). In each arm, a subset of patients was further randomized to undergo tandem colonoscopy; CADe followed by CC or CC followed by CADe. Primary objective was adenoma per colonoscopy (APC). Secondary objectives were adenoma detection rate (ADR) and adenoma miss rate (AMR). Outcomes were also evaluated by colonoscopy indication (screening and surveillance), adenoma location, and adenoma size.

Results: In total, 950 patients were enrolled, of which 916 completed the assigned colonoscopy, 449 in the CADe-assisted group and 467 in the CC group. APC was higher in CADe-arm compared to CC (0.70 vs. 0.51, p=0.015; total adenomas, 314 vs. 238). Overall, ADR was higher in CADe compared to CC (37% vs. 30%, p=0.014). Apart from diminutive (0-5mm) adenomas, use of CADe also increased the detection of small (6-9mm) adenomas compared to CC (14.3% vs. 9.9%, p=0.036). Moreover, an increase in proximal adenoma detection was observed in CADe-assisted colonoscopy compared to CC (46.6% vs. 31.1%, p=0.006). A total of 127 (61 CADe first, 64 CC first) patients completed tandem colonoscopy. AMR was 19% in CADe first compared to 36% in CC first (p=0.024). Use of ME-CADe had no impact on withdrawal times (p=0.861).

Conclusion: ME-APDS increased adenoma detection (both APC and ADR) in non-iFOBT screening and surveillance colonoscopies, and reduced AMR by two-fold compared to CC. Apart from diminutive lesions, ME-APDS increased the detection of 6-9mm adenomas suggesting that this novel CADe system is also able to detect more clinically relevant lesions.

Background:
Artificial intelligence (AI) could revolutionize endoscopy. The first US-approved computer-aided detection (CADe) device (GI Genius) significantly increased adenoma detection rate (ADR) and adenomas per colonoscopy (APC) in randomized controlled trials (RCTs). We sought to replicate the RCT findings in a pragmatic implementation trial.

Methods:
We evaluated CADe for 3 months in a pragmatic trial that leveraged our Colonoscopy Quality Assurance Program and performed a difference-in-difference analysis on quality metrics (with 95% CI) during “Implementation” vs “Pre-implementation” periods in the endoscopy site equipped with CADe (CADe site) vs. 5 non-CADe Control sites, accounting for age, sex, colonoscopy indication and within-endoscopist correlation. Endoscopists at Control sites were not made aware about the CADe trial, and we used a minimalist deployment strategy (no specific expectations/encouragement).

Results:
CADe was used in 1,008/1,037 (97.2%) eligible colonoscopies, with 619 done for screening/surveillance by 24 endoscopists. Study cohort demographics and indications were comparable across sites and periods.

In the CADe site, ADR was 40.1% (36.2-44.0%) and mean APC was 0.78 (0.68-0.90) during Implementation vs 41.8% (37.9-45.8%)(p=0.44) and 0.89 (0.77-1.02)(p=0.23) during Pre-Implementation without CADe (FigA,B). Sessile serrated lesion (SSL), advanced adenoma/SSL, and lesion multiplicity detection rates were also comparable across periods (Table). In Control sites, Implementation vs Pre-Implementation results were comparable (FigA,B; Table).

No effect of CADe on ADR (OR 1.14 [0.83-1.56], p=0.41), APC (OR 1.08 [0.80-1.45], p=0.63) or any other metric was detected (FigA-E, Table). No effects of CADe on procedure times and non-neoplastic lesion resection were seen (Table).

CADe use did not mitigate differences in ADR or APC (FigD,E), or any other metric between lower vs. higher tertiles of performance.

Discussion:
Our results contrast sharply with those of RCTs. Why?

We were interested in real-world CADe implementation, and thus we simply deployed CADe with basic start-up training but no attempt to influence performance or discuss hypotheses. Endoscopists may have dismissed adenomas/SSLs not highlighted by CADe, or true positive CADe prompts. Most concerning would be if CADe led to an unconscious degradation in the quality of mucosal exposure (false sense of comfort?).

In contrast, RCT endoscopists knew the study design/hypotheses and could not be blinded, and they must have known they could influence use of a nascent technology (we do not in any way imply conscious bias or inappropriate conduct).

We remain optimistic about AI, but a minimalist deployment strategy may not ensure success. Better understanding of subtle AI/endoscopist interactions (e.g. mucosal exposure) could promote uniformly high quality in endoscopy.

Background
Insufficient endoscopist experience was independently associated with a lower adenoma detection rate (ADR). The effect of real-time computer-aided polyp detection systems (CADe) on inexperienced endoscopists-in-training during colonoscopies remains unknown.

Method
We performed a single-blind, parallel-group, randomized controlled trial in Hong Kong during April 2021 and July 2022. Consecutive eligible subjects undergoing colonoscopies for screening/surveillance/diagnostic purposes were randomized in 1:1 ratio to receive colonoscopies with the use of real-time CADe system (ENDOAID, Olympus) or without it (control) during the withdrawal phases. (Figure 1) All colonoscopies were performed by endoscopists-in-training with a personal experience of <500 procedures and less than 3 years. Randomization was stratified by age, gender and endoscopist experience (beginner vs intermediate level, <200 vs 200-500 procedures). Image enhancement or device assistance were not allowed. Subjects with incomplete colonoscopies or low Boston Bowel Preparation Scale (BBPS) were excluded. Treatment allocation was blinded to outcome accessors (pathologists) but not endoscopists. The primary outcome was ADR and secondary outcomes included ADR of different sizes and sites, mean number of adenomas and non-neoplastic resection rate.

Results
Among 766 subjects included, 386 and 380 subjects were randomized to CADe and control groups respectively. Baseline characteristics were comparable. The overall ADR was significantly higher in CADe group than control group (57.0% vs 43.9%, p<0.001). (Table 1) The ADRs for <5mm (40.2% vs 24.5%, p<0.001) and 5-10mm adenomas (36.5% vs 28.7%, p=0.025) were significant higher in CADe group. The ADRs were significantly higher in CADe group at both right-sided (41.7% vs 30.3%, p=0.001) and left-sided colon (33.9% vs 27.1%, p=0.048). The ADRs were significantly higher among beginner (57.9% vs 39.3%, p=0.008) and intermediate level endoscopists (56.6% vs 45.8%, p=0.015) in CADe group. Mean number of adenomas per colonoscopy was significantly higher in CADe group (1.46 vs 0.85, p<0.001). Non-neoplastic resection rate was higher in CADe group (51.8% vs 34.5%, p<0.001). There was no significant difference for advanced ADR (7.3% vs 8.4%) and sessile serrated lesion detection rate (2.1% vs 1.8%). A multivariate analysis demonstrated that advanced age, male gender, total withdrawal time, endoscopist experience and the use of CADe were significant factors for a higher ADR. The use of CADe remained an independent factor for a higher ADR after adjustment by logistic regression model (odds ratio 1.56, 95%CI 1.13-2.18, p=0.008).

Conclusion
The use of real-time CADe during colonoscopies in endoscopists-in-training could increase the overall ADR, especially for small-to-medium size adenomas, in different levels of experience. (NCT04838951)

Figure 1. Study Flow Diagram

<b>Table 1</b>. Baseline characteristics and outcomes between real-time computer-aided polyp detection systems (CADe) and control groups. SD: standard deviation. ADR: adenoma detection rate. p-values were calculated by Pearson Chi-square test or Wilcoxon Rank-Sum test where appropriate.

Table 1. Baseline characteristics and outcomes between real-time computer-aided polyp detection systems (CADe) and control groups. SD: standard deviation. ADR: adenoma detection rate. p-values were calculated by Pearson Chi-square test or Wilcoxon Rank-Sum test where appropriate.

Background: Colonoscopy, the gold standard in colon cancer screening (CCS), is an imperfect screening tool due in part to differences in operators and technical skills, leading to an adenoma missed rate as high as 26%. Adenoma detection rate (ADR) is an important quality metric in CCS, and several interventions have been developed to increase ADR, including computer-aided detection (CADe) technology and endocuff, which have been shown to increase ADR. Despite this potential benefit, CADe and endocuff have yet to be widely adopted, with a paucity of data regarding the use of CADe in conjunction with other tools such as endocuff.

Methods: In this prospective quality improvement (QI) initiative, adults 18 years or older undergoing screening, surveillance, or diagnostic colonoscopy utilizing real-time AI technology (Medtronic GI GeniusTM) at a community hospital and ambulatory endoscopy center were analyzed. Diagnostic exams included evaluations of GI bleed, anemia, and IBD. Exclusion criteria included standard procedural contraindications such as acute diverticulitis, known/suspected perforation, a Boston Bowel Preparation Scale (BBPS) score of <6 or segment score of 0-1, and incomplete examinations. ADR and adenoma per colonoscopy (APC) were calculated with the addition of CADe versus endoscopist alone.

Results: 147 patients underwent both inpatient and outpatient colonoscopy exams, 134 of which were eligible based on the inclusion criteria. The overall ADR was 42.9% (58/134), with no statistically significant difference in ADR when comparing endoscopist alone (41.0%) vs endoscopist with CADe assistance (44.8%), p=0.6216. CADe assistance resulted in a surplus of 10.4% polyps per colonoscopy (PPC), however, APC did not increase significantly (5.97%), p=0.2975. CADe assistance along with endocuff was utilized in 40.7% (54/134) of cases. The use of CADe plus endocuff resulted in an ADR of 61.1% (33/54) vs 31.25% (25/80) in CADe without endocuff, p-value = 0.0008. APC was also significantly increased in CADe plus endocuff 125.9% (68/54) vs CADe without endocuff 51.25% (41/80).

Conclusions: Our results indicate that endoscopists alone were able to identify a comparable number of polyps and adenomas with or without CADe assistance. Notably, the application of CADe technology in conjunction with endocuff resulted in a significant increase in ADR and APC, suggesting that tissue exposure in addition to visualization, is of great importance during the procedure. Additional studies are required to further validate these results in the inpatient and outpatient settings, as well as the potential synergistic effect of CADe plus endocuff.

Adenoma Detection Rate (ADR)

Background: Complete polyp resection is critical for colorectal cancer prevention. Polyp boundaries should be clearly identified to assist resection. Computer-aided image segmentation is an image processing task that can identify meaningful structures and polyp margins, which can be performed with supervised machine learning methods by using both images and segmentations for training. Convolutional neural network (CNN) based methods represent the current state-of-the-art. However, these approaches require a high number of expert-labeled images, which is costly and time intensive. Unsupervised machine learning approaches have promise in identifying meaningful structures and edges without human annotation. We present a novel contrastive unsupervised learning framework for polyp segmentation in colonic polyps to automatically identify polyp outlines and compare its performance to supervised algorithms.

Methods: Our unsupervised learning approach is based on a contrastive learning framework utilizing CNN. The algorithm is designed for deployment in conjunction with a polyp detection algorithm. After extracting image patches, a structural similarity loss is calculated on a pixel-centered-patch level basis to find intra-image contrast boundaries, and then similar/dissimilar patch pairs are fed into a CNN to learn latent representations. This was trained on a dataset of paired white light (WL) and narrow band images (NBI) of colonic polyps by using two modalities as the dual-channel input of the model, and externally validated on the publicly available Kvasir polyp dataset (WL only). Segmentation performance was evaluated using the Dice coefficient.

Results: The contrastive unsupervised learning approach trained on WL images achieved a Dice (mean±SD) of 0.71±0.08 on internal validation and 0.69±0.17 on external validation. Compared to the supervised method, which has a Dice of 0.90±0.08 and 0.84±0.16, our approach shows a smaller decay in performance, which suggests better generalizability. A combination of NBI images and WL images show a slight improvement in performance with a coefficient of 0.72±0.10 compared to training on WL or NBI images alone.

Conclusion: An unsupervised method for polyp segmentation maintains performance on external validation, though still does not outperform supervised neural networks.

Harry Aslanian

Appearances