Use of Artificial Intelligence Image Classifier for Real-time Detection of Colonic Polyps

Abstract

Background: Artificial intelligence (AI) had been shown to be potentially useful in colonic polyp detection. However, its performance under special imaging methods e.g. narrow band imaging (NBI) had not been studied. One of the major challenge of AI is to detect small polyps (<5mm) because it is sometimes also difficult for usual endoscopist.

Methods: An AI-NBI colonic polyp detector which was built on a convolutional neural network (with 5 convolutional layers and 3 fully connected layers) was trained by 8500 NBI endoscopic images from both normal colon and colonic image with polyp(s). Six independent real-time colonoscopy videos were prospectively collected for validation and all of them are withdraw under NBI from caecum. The minimal withdrawal time for each validation colonoscopy video was 6 minutes. The AI-NBI colonic polyp detector analyzed all frames from each colonoscopy video including those frames with blurred images starting from the withdrawal from caecum. The AI-NBI colonic polyp detector would signal polyp detection when the average probability of colonic polyp detection from 50 video frames (2 seconds of video) was more than 90%. A correct polyp detection by AI-NBI colonic polyp detector is defined by any polyp detection signal is presence when 2 seconds video being analyzed containing a polyp detected by the original endoscopist. And all colonoscopy was performed by endoscopists with at least 2000 NBI colonoscopy experience. The endoscopies used in this study were non –optical magnifying endoscopy series (CF-H260, CF-HQ290 model) and CV-260 or CV-290 video system (Olympus Medical system) were used.

Results: The AI colonic polyp detector can identify all 20 colonic polyps that are detected by endoscopists with mean size of 2.6mm. In term of morphology, most of them are Paris type 0 IIA (90%, n=18) and two of them are Is (10%, n=2). More than half of them (60%, n=12) are hyperplastic polyps and remaining polyps are tubular adenoma (40%, n=8). The average bowel preparation score (Aronchick scale 3) for all colonoscopic videos was 4.9. A total of 152095 frames from 6 video were analyzed. In term of frame analyzing performance of AI-NBI colonic polyp detector, it had the sensitivity 98.3% (95%CI: 98.0% - 98.6%), specificity 99.7% (95%CI: 99.6% -99.7%), positive predictive value 95.6% (95%CI: 95.2%- 96.0%) and negative predictive value 99.9%. (95%CI: 99.8% - 99.9%). The accuracy was 99.6% (99.6%- 99.6%) The area under ROC curve was 0.99 (95%CI: 0.98 – 0.99). The major reason of false positive signal was a “suction- polyp”. The major reason of false negative signal was a large series of blurred frames during that period of video frame analysis.

Conclusions: AI-NBI colonic polyp detector is feasible for real-time detection of colonic polyp with high accuracy. Further studies are needed to confirm its role in clinical practice.