Deep learning methods for automated orthopaedic measurements in radiographs

Abstract

Radiography remains central to orthopedic diagnosis and surgical planning, yet many clinically important measurements (angles, offsets, alignment indices, and grading scores) are time-consuming to obtain and show intra- and inter-observer variability, especially in early or borderline disease. A reader-friendly overview of deep learning methods for the automation of orthopedic measurements on radiographs is provided in this article. Four common approaches are summarized: segmentation networks through which masks are generated and geometry-based measurements are derived; key point or heatmap-based landmark detection models through which automated angle and distance calculation is enabled; two-step object detection plus classification pipelines in which regions of interest are localized before downstream prediction; and direct regression models through which measurements are output directly from images. Representative applications are highlighted across major regions, including hip dysplasia angles, knee alignment and osteoarthritis assessment, and shoulder measurements such as the critical shoulder angle. Practical limitations and implementation considerations are also discussed, including generalizability across sites, performance drift, bias and fairness, and the need for human oversight and interpretable outputs to support safe deployment.