Artifical Intelligence Image Classifier Based on Nonoptical Magnified Images Accurately Predicts Histology and Endoscopic Resectability of Different Colonic Lesions

Abstract

Background: We evaluated the use of artificial intelligence (AI) assisted image classifier, which was based on non-optical magnified endoscopic narrow band images (NBI), in predicting the histology of solitary colonic lesions. Methods: An AI image classifier which was built on a convolutional neural network (with 5 convolutional layers and 3 fully connected layers) was first trained by 8,748 endoscopic images from 3,509 colonic lesions including 2,715 tubular/tubulovillous adenoma, 501 hyperplastic polyps/serrated polyps and 293 adenocarcinoma. The independent validation set consisted of 345 regions of interest (ROI) within the endoscopic images (which included all endoscopically visible abnormal mucosal area) highlighted from another 126 colonic lesions. All endoscopic images used in this study were obtained by non–optical magnifying colonoscopy series (CF-H260, CF-HQ290 model; Olympus Medical System) with NBI function. The histology of these lesions was based on the most severe histology noted in the resected specimen of the whole lesion or in multiple biopsies. The prediction from the trained AI image classifier on these ROI images was compared with the final histology (Figure). Results: The mean size of the 126 colonic lesions in the validation set was 13.7 mm (range: 1 to 60 mm). The locations of these lesions were in caecum (6.3%), ascending colon (20.6%), transverse colon (15.1%), descending colon (11.1%), sigmoid (42.1%) and rectum (4.8%). The most common histological types were tubular adenoma (47.6%), followed by carcinoma with deep invasion (15.9%), carcinoma with superficial invasion (7.9%) hyperplastic polyps (14.3%), sessile serrated polyps (7.9%) and tubulovillous adenoma (6.6%). The performance of the AI image classifier for prediction of different histological types is summarized in Table. The sensitivity of hyperplastic/serrated polyps was 96.6% and NPV was 99.6%, but was lower in tubular adenoma and cancer. If only diminutive (<5mm) colonic polyps were analyzed, the correlation of surveillance colonoscopy interval using AI image classifier and histology was 0.97 (p <0.0001). The AI image classifier also had a high accuracy (88.2%) in prediction of carcinoma with deep invasion, which is not endoscopically curable. Conclusions: The trained AI image classifier can accurately predict the histology of a whole spectrum of colonic lesions. For diminutive polyps, the trained AI met the ASGE PIVI threshold of 90% in the assignment of post-polypectomy surveillance intervals. High negative predictive value and accuracy of the AI image classifier for carcinoma with deep invasion also suggested that it can help to select curable lesion for endoscopic removal