Identification of Pain-Associated Effusion-Synovitis from Knee Magnetic Resonance Imaging by Deep Generative Networks

ABSTRACT Objectives To identify the source and location of osteoarthritis-induced pain symptoms, we used deep learning techniques to identify imaging abnormalities associated with pain from magnetic resonance imaging (MRI) of knees with symptoms of symptoms of osteoarthritis pain. Methods Pain-associated areas were detected from the difference between the MRI images of symptomatic knees and their respective counterfactual asymptomatic images generated by a Generative adversarial network. A total of 2,225 pairs of 3D MRI images were extracted from patients with unilateral pain symptoms in the baseline and follow-up cohorts of the Osteoarthritis Initiative. Subsequently, pain-associated effusion-synovitis were characterized into subregions (patellar, central, and posterior) using an anatomical segmentation model. Results We found that the volumes of pain-associated effusion-synovitis were more sensitive and reliable indicators of pain symptoms than the overall volumes in the central and posterior subregions (odds ratio [OR]:3.23 versus 1.77 in the central region, and 3.18 versus 2.66 in the posterior region for severe effusion-synovitis). For mild effusion-synovitis, only pain-associated volume was found to be associated with pain symptoms, but not with overall volume. Patients with significant pain-associated effusion-synovitis in the patellar subregion had the highest increased odds of pain symptoms (OR=4.86). Conclusion To the best of our knowledge, this is the first study to utilize deep-learning-based models for the detection and characterization of pain-associated imaging abnormalities. The developed algorithm can help identifying the source and location of pain symptoms and in designing targeted and individualized treatment regimens.