Foundation Models Meet Musculoskeletal Digital Twins: Uncertainty-Aware Early Detection of Subclinical Cartilage...

Detecting early cartilage degeneration on MRI remains challenging because early compositional changes produce only subtle signal variations easily confounded with normal anatomical and acquisition variability. We pro pose CartiDT, a cross-modal transfer learning framework that couples vision foundation models with musculoskeletal digital twin representations for preclinical cartilage dam age detection. CartiDT comprises three modules: (1) a parameter-efficient adaptation module that transfers frozen vision transformer features to three-dimensional double echo steady-state knee MRI via depth-wise Low-RankAdap tation and depth-axis positional interpolation; (2) a graph neural network encoder whose biomechanically derived adjacency captures inter-compartment dependencies within a lightweight digital twin; and (3) an evidential deep learning head that decomposes predictive uncertainty into aleatoric and epistemic components, enabling the model to abstain on unreliable predictions. Evaluated on the Osteoarthri tis Initiative longitudinal dataset (4,796 subjects), CartiDT achieved an area under the receiver operating characteristic curve of 0.861 (±0.003) for preclinical damage detection and a Dice similarity coefficient of 0.879 (±0.002) for six-compartment segmentation, surpassing the strongest baselines by 3.8 and 1.2 percentage points, respectively. Selective prediction based on uncertainty awareness eliminated 29.6% of false positives when compared to the most powerful uncertainty-aware baseline (the equivalent of a five-member deep ensemble) with a 15% referral bar, with only one forward pass (3.2 s/volume, compared to 16 s/volume); the increased calibration and the increase in speed are positive and separate benefits. Patellar cartilage remained the most difficult compartment (Dice 0.814, detection AUROC 0.806), and the digital twin module added a modest 18% wall-clock overhead versus segmentation alone. These results demonstrate that integrating foundation model representations with structured anatomical priors enhances early cartilage damage detection, while thin-cartilage regions and computational cost remain open challenges.