Performance of artificial intelligence system for interrogation of scars after piecemeal endoscopic mucosal resection of large non pedunculated polyps

Aims The detection and management of recurrence after piecemeal endoscopic mucosal resection (p-EMR) of large colorectal polyps are crucial for preventing post-colonoscopy colorectal cancer. Artificial intelligence systems for lesion detection during colonoscopy (CADe) have shown efficacy in identifying small lesions (< 10 mm). Our primary aim was to evaluate the performance of CADe compared to endoscopists in detecting recurrence after p-EMR. Secondary objectives included comparing the performance of endoscopists based on their expertise (experts vs. non-experts), assessing CADe and endoscopist performance according to histopathology categories (adenomas vs. serrated lesions) and evaluating interobserver agreement.

Methods A total of 66 anonymized, high-quality videos of post-EMR scars from polyps>15 mm were evaluated, including 15 with previously clipped defects. Histopathological recurrence was present in 25 out of 66 cases. Six endoscopists (three experts and three non-experts) predicted the presence of recurrence (yes/no) and assigned a confidence level (high/low). The same videos were analyzed using CAD-EYE (ELUXEO 7000; Fujifilm) with the detection mode activated. Histopathology was the gold standard. All measures were expressed as percentages along with 95% confidence intervals based on Wilson method. Comparison between AI and endoscopists performance were analyzed with Mc Nemar test. Observer’s concordance pairwise agreement were tested with Fleiss Kappa.

Results Recurrence size was<10 mm in 19 out of 25 cases (76.0%). Endoscopists demonstrated higher sensitivity than CADe in detecting recurrence: 96.0% [95% CI 80.5-99.3] vs. 72.0% [95% CI 52.4-85.7]; p=0.03. Non-expert endoscopists also exhibited better sensitivity compared to CADe: 92.0% [75.0-97.8] vs. 72.0% [95% CI 52.4-85.7]. For non-clipped defects, endoscopists maintained higher sensitivity at 95.0% [95% CI 76.4-99.1] compared to CADe at 65.0% [95% CI 43.3-81.8]; p=0.03. There were no significant differences in performance between CADe and endoscopists according to histopathology categories (adenoma vs. serrated lesions). Interobserver agreement among endoscopists was moderate for the entire cohort (Kappa 0.57; [95% CI 0.49-0.61]), substantial among experts (Kappa 0.70; [95% CI 0.56-0.85]), and increased when excluding previously clipped defects (Kappa 0.61; [95% CI 0.54-0.68]). Experts demonstrated better performance in ruling out recurrence compared to non-experts, with a specificity of 92.7% [95% CI 80.6-97.5] versus 58.5% [95% CI 43.4-72.2]; p=0.001, and accuracy of 92.4% [95% CI 83.5-96.7] versus 72.7% [61.0-82.0]; p<0.001.

Conclusions Endoscopists outperformed CADe in evaluating post-EMR scars, even among non-experts and after excluding clipped defects. Training CADe with images of post-polypectomy scars could further enhance its performance.

Publikationsverlauf

Artikel online veröffentlicht:
27. März 2025