Validation of YOLOv8 algorithm in detecting colon polyps in endoscopy videos

<p dir="ltr">Background: Colorectal cancer is the fifth most common cancer in Vietnam, with rapidly increasingmortality. Detecting polyps, especially adenomas in colonoscopy, plays an important role in reducing colorectal cancer mortality. Artificial intelligence (AI) shows promising results in increasing detection rate of colorectal polyps and adenomas; however, it has not been tested on endoscopic data of patients from Vietnam. YOLOv8-integrated endoscopy with enhanced accuracy, low latency has been developed by Vietnamese experts, but it has not been validated on endoscopic videos. Therefore, our research aims to (I) evaluate YOLOv8 algorithm’s accuracy in colorectal polyp detection on endoscopic videos and (II) describe common false detections and misses made by YOLOv8 algorithm in colorectal polyp detection.</p><p dir="ltr">Methods: A cross-sectional study was conducted on a colonoscopy dataset of 50 videos and 20,616 images. In which, the videos, each with a minimum withdrawal time of 5 minutes, were curated at Institute of Gastroenterology and Hepatology from December 2022 to June 2023, the endoscopic images were collected from Hanoi Medical University Hospital and Institute of Gastroenterology and Hepatology. The YOLOv8 algorithm was trained on the images dataset and 20 colonoscopy videos labeled by experts. The algorithm was subsequently validated on 30 remaining colonoscopy videos by comparing AI’s segmentations with the ground-truth (labeled by experts). The accuracy was assessed using recall, precision, and F1-score.</p><p dir="ltr">Results: Thirty validation videos comprised of 69,003 frames with polyps and 399,818 without polyps. A total of 68 polyps were identified, classified into only 3 morphologies Ip, Is, and IIa (based on Paris classification). The YOLOv8 algorithm, with a processing speed of 72 frames per second (fps), detected 65 out of 68 polyps (95.6%). We compare all the metrics of two IoU settings (0.3 and 0.5) at a fixed confidence score of 0.5. With intersection over union (IoU) =0.5, recall, precision, and F1-score of the algorithm were 0.695 [95% confidence interval (CI): 0.691, 0.698], 0.482 (95% CI: 0.479, 0.485), and 0.57 (95% CI: 0.566, 0.572), respectively. Lowering the IoU setting to 0.3 improves the algorithm’s recall (from 0.69 to 0.74). False positives were mainly due to enteric mucosal folds (42.5%), endoscopic instruments (17.3%), and polyp remnants after resection (12.2%). Most false negatives were polyps that have vague surfaces (45.2%), diminutive or distant (21.4%), or obscured (15.7%).</p>