Reverse-engineering the diagnostic process: An indirect finding-driven deep learning model for pelvic active bleeding detection on CT that achieves specialist-level performance and surpasses residents — a multicenter study