Identification of patients with malignant biliary strictures using a cholangioscopy-based deep learning artificial intelligence (with video).

BACKGROUND AND AIMS: Accurately diagnosing malignant biliary strictures (MBSs) as benign or malignant remains challenging. It has been suggested that direct visualization and interpretation of cholangioscopy images provide greater accuracy for stricture classification than current sampling techniques (ie, brush cytology and forceps biopsy sampling) using ERCP. We aimed to develop a convolutional neural network (CNN) model capable of accurate stricture classification and real-time evaluation based solely on cholangioscopy image analysis. METHODS: Consecutive patients with cholangioscopy examinations from 2012 to 2021 were reviewed. A CNN was developed and tested using cholangioscopy images with direct expert annotations. The CNN was then applied to a multicenter, reserved test set of cholangioscopy videos. CNN performance was then directly compared with that of ERCP sampling techniques. Occlusion block heatmap analyses were used to evaluate and rank cholangioscopy features associated with MBSs. RESULTS: One hundred fifty-four patients with available cholangioscopy examinations were included in the study. The final image database comprised 2,388,439 still images. The CNN demonstrated good performance when tasked with mimicking expert annotations of high-quality malignant images (area under the receiver-operating characteristic curve, .941). Overall accuracy of CNN-based video analysis (.906) was significantly greater than that of brush cytology (.625, P = .04) or forceps biopsy sampling (.609, P = .03). Occlusion block heatmap analysis demonstrated that the most frequent image feature for an MBS was the presence of frond-like mucosa/papillary projections. CONCLUSIONS: This study demonstrates that a CNN developed using cholangioscopy data alone has greater accuracy for biliary stricture classification than traditional ERCP-based sampling techniques.