Standard knee radiographs enable deep learning inference of MRI-defined cartilage and meniscal damage in early knee osteoarthritis: a study using the osteoarthritis initiative database.

BACKGROUND

To develop and validate a radiograph-only deep learning framework that jointly predicts magnetic resonance imaging (MRI)-defined early structural pathology in knees with early radiographic osteoarthritis (Kellgren-Lawrence [KL] grades 0-1) and performs full-spectrum KL classification across all grades within a single multi-task architecture.

METHODS

This retrospective study used baseline data from the Osteoarthritis Initiative (OAI) comprising 8260 knees (4130 participants). A multi-task ConvNeXt-Base network was trained with five-fold stratified group cross-validation enforcing strict subject-level data separation. KL classification was supervised across all knees. MRI prediction heads were activated only for KL 0-1 knees with available labels. MRI Osteoarthritis Knee Score (MOAKS)-derived binary endpoints were defined at a threshold of ≥2: tibiofemoral cartilage damage (primary endpoint) and meniscal morphology damage (secondary endpoint).

RESULTS

KL grading performance across held-out folds was as follows: mean quadratic weighted kappa (QWK) 0.8284 (standard deviation [SD] 0.0255), balanced accuracy 0.6836 (SD 0.0159). In KL0/1 knees with both MRI labels (n = 2561), prevalence was 31.3% for cartilage damage and 26.4% for meniscal damage. Mean AUROC/AUPRC were 0.7329 (SD 0.0203)/0.5741 (SD 0.0364) for cartilage damage and 0.7193 (SD 0.0533)/0.5147 (SD 0.0719) for meniscal damage.

CONCLUSIONS

A leakage-controlled multi-task radiograph model achieved strong KL grading agreement and moderate discrimination of MRI-defined cartilage and meniscal pathology in knees with early radiographic osteoarthritis, supporting a potential role as an assistive triage signal to identify patients who may benefit from earlier MRI evaluation. However, clinical correlation remains essential; MRI referral decisions should always be made in the context of the individual patient's symptoms, functional impairment, and overall clinical assessment, and cannot be based on AI-derived probability scores alone.