Automatic Measurement of Joint Space Width from Hand Radiographs Using Deep Learning Models

Joint Space Width (JSW) is a measurement to diagnose hand osteoarthritis from radiographs. This study proposed an automatic method to measure JSW using a modified VGG19 model and U-net segmentation as the pre-processing step. With a dataset of 3591 hand radiographs, 34,597 DIP/PIP/MCP joints were cut. The ground truth was labeled by healthcare professionals. We first trained individual models for each type of the DIP, PIP, and MCP joints and finally trained a combined model for all. The individual models produced correlation coefficients of 0.88, 0.88, and 0.90 and mean square error (MSE) of 0.02 mm, 0.02 mm, and 0.03 mm for DIP, PIP, and MCP joints, respectively. The combined model further improved the performance to 0.89, 0.89, and 0.98 in correlation and the same MSE. The result demonstrates that the proposed method is robust to different sizes and shapes of hand joints and can save the time of manual labeling.