DEEP‐LEARNING ASSESSED BONE MARROW LESIONS CAN PREDICT FUTURE RADIOGRAPHIC OSTEOARTHRITIS INCIDENCE AND SYMPTOM WORSENING: OAI DATA

Abstract

Bone marrow lesions (BMLs) are the MRI hallmark of bone involvement in knee OA and correlate with knee pain. While they may have great prognostic values, their assessment in large-scale datasets can be time-consuming and subject to inter-reader variability. we aimed to develop a validated deep-learning (DL) method for accurate classification of knee BMLs and to investigate their predictive role in the radiographic incidence and symptom worsening in patients with baseline pre-radiographic knee OA (i.e., KLG <2). We used coronal MPR and sagittal intermediate-weighted MRIs from 2488 knees with available semi-quantitative MRI Osteoarthritis Knee Scores (MOAKS) reading of the Osteoarthritis Initiative (OAI) cohort to train and 5-fold cross-validation of convolutional neural networks. Using the developed DL algorithm, participants were dichotomized according to baseline BML status as “no/minimal” (≤2 knee subregions affected and maximum BML grade ≤1) or “moderate/severe.” The model was then used to predict BML status at the baseline visit of the remaining knees without available MOAKS readings. Knees with moderate/severe and with no/minimal BMLs were selected using 1:1/2 propensity score (PS) matching for potential confounders. Survival and mixed models were used to assess the association between baseline BML and risk and odds of radiographic incidence and symptom worsening over 9 years, respectively. The DL model had an AUC of 0.84±0.01, sensitivity of 0.85±0.02, and specificity of 0.63±0.03. The PS matched cohorts with moderate/severe and no/minimal BML consisted of 2065 and 1236 knees. Having moderate/severe BMLs at baseline was associated with almost twice the risk of knee OA radiographic incidence (hazard ratio, 95%CI: 1.92, 1.56–2.36) and odds of symptom worsening (WOMAC total score OR:1.76, 1.07–2.88) in 9 years of follow-up in knees with pre-radiographic knee OA. DL models can classify knee MRIs according to BML status with combined MOAKS measures of the number of affected subregions and maximum BML grade. Using these predictions, we could show that baseline BMLs can predict downstream knee OA incidence and symptom worsening in patients with pre-radiographic knee OA. NIH National Institute of Aging (NIA) under Award Number P01AG066603 and NIH National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS) under Award Number R01AR079620-01. AG reported receiving funding from Merck Serono, AstraZeneca, Galapagos, Pfizer, Roche, TissueGene (for consultation), and Boston Imaging Core Lab (as the president and stockholder). SD reported that he received funding from Toshiba Medical Systems (for consultation) and grants from GERRAF and Carestream Health (for a clinical trial study). Authors acknowledge OAI staff and team. CORRESPONDENCE ADDRESS: sdemehri1@jhmi.edu .