A novel deep learning network for small bowel ulcerative lesion detection and differential diagnosis on double-balloon endoscopy images.

Background And Aims.Differentiating small bowel ulcerative diseases (SBUDs) on double-balloon enteroscopy(DBE) is challenging. We aimed to develop an artificial intelligence model using DBE images for accurate SBUD identification and classification.Methods.We retrospectively analyzed 1791 double-balloon endoscopy (DBE) images from 283 patients diagnosed with five types of SBUDs at a single center from August 2020 to May 2023. The cohort included Crohn's disease (n= 187), cryptogenic multifocal ulcerous stenosing enteritis (CMUSE) (n= 37), intestinal tuberculosis (n= 15), non-specific ulcer (n= 31), and primary small intestinal lymphoma (n= 13). Ulcerative lesions were delineated by three endoscopists and finalized by consensus from at least two experts. A novel cascade network, Cascade-E-Yolov7 (EfficientNet-B1 and ESFC-Yolov7), was developed for the classification and precise localization of these lesions.Results.For classification, the model achieved an overall accuracy of 82.35%. The area under the curve was 0.90 for CD, 0.96 for CMUSE, 0.83 for ITB, 0.91 for non-specific ulcers, and 0.95 for PSIL. For lesion detection, the model yielded a mean average precision (mAP@0.5) of 81.35%, a precision of 82.15%, and a recall of 72.67%.Conclusions.The Cascade-E-Yolov7 model accurately detected and classified SBUDs, showing potential as a clinical tool to facilitate diagnosis and reduce experience-dependent variability.