A Deep Learning Tool for Minimum Joint Space Width Calculation on Antero-posterior Knee Radiographs

Background: Minimum joint space width (mJSW) is an important continuous quantitative metric of osteoarthritis progression in the knee. The purpose of this study was to develop an automated measurement algorithm for mJSW in the medial and lateral compartments of the knee that can flexibly handle native knees and knees after arthroplasty. Methods: We developed an end-to-end algorithm consisting of a deep learning segmentation model plus a computer vision algorithm to measure mJSW in the medial and lateral compartments of the knee. Trained annotators segmented 583 images to train, validate, and test a deep learning model that segments the relevant structures for the measurement of mJSW. Trained annotators measured mJSW in 330 independent images to provide ground truth measurements for the development and validation of the computer vision algorithm. Algorithm performance was measured by calculating mean absolute error and constructing the Bland-Altman plots. Results: The trained segmentation model performed with an average dice score of 0.92 across all images and structures in the 50-image test set. The mean absolute error between the human measurements and the algorithm measurements was 0.85 ± 1.20 mm. The mean error without taking the absolute value was 0.019 mm, demonstrating minimal bias toward overestimating or underestimating mJSW. Of the algorithm mJSW measurements, 73.2% were less than 1 mm and different from human measurements. Conclusions: We developed and validated an automated algorithm for measuring mJSW on anteroposterior knee radiographs. This artificial intelligence–based algorithm for mJSW will streamline future population-level clinical research in the natural history of the knee joint and allow surgeons and physicians the ability to quickly measure a patient's entire longitudinal series of radiographs to quantitatively assess the radiographic progression of arthritis during clinic visits.