Artificial Intelligence and Polyp Detection in Colonoscopy: Use of a Single Neural Network to Achieve Rapid Polyp Localization for Clinical Use.

Abstract

BACKGROUND AND AIMS\nArtificial intelligence has been extensively studied to assist clinicians in polyp detection, but such systems usually require expansive processing power, making them prohibitively expensive and hindering wide adaption. The current study used a fast object detection algorithm, known as the YOLOv3 algorithm, to achieve real-time polyp detection on a laptop. In addition, we evaluated and classified the causes of false detections to further improve accuracy.\n\n\nMETHODS\nThe YOLOv3 algorithm was trained and validated with 6,038 and 2,571 polyp images, respectively. Videos from live colonoscopies in a tertiary centre and those obtained from public databases were used for the training and validation sets. The algorithm was tested on 10 unseen videos from the CVC-Video ClinicDB dataset. Only bounding boxes with an intersection over union area of > 0.3 were considered positive predictions.\n\n\nRESULTS\nPolyp detection rate in our study was 100%, with the algorithm able to detect every polyp in each video. Sensitivity, specificity and F1 score were 74.1%, 85.1% and 83.3, respectively. The algorithm achieved a speed of 61.2 frames per second (fps) on a desktop RTX2070 GPU and 27.2 fps on a laptop GTX2060 GPU. Nearly a quarter of false negatives happened when the polyps were at the corner of an image. Image blurriness accounted for approximately 3% and 9% of false positive and false negative detections, respectively.\n\n\nCONCLUSION\nThe YOLOv3 algorithm can achieve real-time poly detection with high accuracy and speed on a desktop GPU, making it low-cost and accessible to most endoscopy centres worldwide.