Fig. 2

Overview of the proposed automatic-bone-destruction-evaluation system. A Input an X-ray image of hands into the detection model (DeepLabCut). DeepLabCut detects the center point of the evaluation joints of the SHS (target joints): 16 joints for erosion (the 4 PIP joints, the IP joint of the thumb, the 5 MCP joints, the CMC joint of the thumb, the multangular, the navicular, the lunate, the radius, and the ulna) and 15 joints for JSN (the 4 PIP joints, the 5 MCP joints, the 3 CMC joints, the multangular-navicular joint, the capitate-navicular-lunate joint, and the radiocarpal joint). Each point indicates the detected center of the target joints. Each joint image was cropped (red bounding box) according to the detected center point. B Each cropped image was input into the classification model, which outputs whether the input image is intact (SHS = 0) or non-intact (SHS ≥ 1). SHS, Sharp/van der Heijde score; PIP, proximal interphalangeal; IP, interphalangeal; MCP, metacarpophalangeal; CMC, carpometacarpal; JSN, joint space narrowing