An AI-powered diagnostic system for grading and invasion of non-muscle-invasive bladder cancer via TURBT specimens: A multicenter study.

Early pathological examination for grading and invasion of non-muscle-invasive bladder cancer (NMIBC) via transurethral resection of bladder tumor (TURBT) specimens is a labor-intensive, subjective, and experience-dependent task, while the poor quality of TURBT specimens and the high heterogeneity of NMIBC tumors further limit diagnostic accuracy and efficiency. This study proposed a dual-channel multi-instance learning (DCMIL) model to simultaneously integrate the grading and invasion of NMIBC, while efficiently locating minor changes in the morphology and distribution of NMIBC cells in parallel. Developed on a multicenter dataset of 1332 whole slide images (WSIs) from TURBT specimens, DCMIL demonstrated outstanding accuracy and robust performance with areas under the curve of 0.851-0.983. In the reader study, DCMIL-assisted interpretation improved the diagnostic accuracy by an average of 13.31%-18.58% for inexperienced pathologists. Overall, DCMIL holds promise as a reliable initial assessment tool of NMIBC via TURBT specimens to support clinical decision-making.