A Deep Learning Tool for Hip Minimum Joint Space Width Calculation on Antero-posterior Pelvis Radiographs.

Abstract

Minimum joint space width (mJSW) is a useful quantitative metric of osteoarthritis progression in the hip, particularly as a continuous variable compared to more common categorical classification systems. The purpose of this study was to develop an automated algorithm for measuring mJSW in native hips on antero-posterior pelvis radiographs. An end-to-end algorithm was developed, consisting of a deep learning segmentation model plus a computer vision algorithm to measure mJSW in the hip joint. Trained researchers annotated 300 radiographs for training and validation of an automated segmentation model that identifies relevant structures for the measurement of mJSW. Trained annotators also independently measured mJSW in 375 additional images to provide ground truth measurements for the development and validation of a computer vision algorithm. External validation was performed on 75 images from the Osteoarthritis Initiative. Algorithm performance was measured by calculating the mean absolute error and constructing a Bland-Altman plot. The mean absolute error between the human and the algorithm's measurements was 0.87 ± 1.05 mm. In 70% of cases, the algorithm's mJSW measurements were less than one mm different from human measurements, in 84% the difference was less than 1.5 mm, and in 90% the difference was less than two mm. In the Osteoarthritis Initiative external validation cohort, mean absolute error was 0.86 ± 0.69 mm. The trained segmentation model obtained an average Dice score of 0.71 across all structures in the test set. An automated model for measuring mJSW on antero-posterior pelvis radiographs was developed and externally validated. This algorithm performs well at the sub-millimeter level and may streamline longitudinal patient evaluation and population-level clinical research in the natural history of the hip joint.