Clinical utility of foundation models in musculoskeletal MRI for biomarker fidelity and predictive outcomes

Precision medicine in musculoskeletal imaging requires scalable measurement infrastructure. We developed a modular system that converts routine MRI into standardized quantitative biomarkers suitable for clinical decision support. Promptable foundation segmenters (SAM, SAM2, MedSAM) were fine-tuned across heterogeneous musculoskeletal datasets and coupled to automated detection for fully automatic prompting. Fine-tuned segmentations yielded clinically reliable measurements with high concordance to expert annotations across cartilage, bone, and soft tissue biomarkers. Using the same measurements, we demonstrate two applications: (i) a three-stage knee triage cascade that reduces verification workload while maintaining sensitivity, and (ii) 48-month landmark models that forecast knee replacement and incident osteoarthritis with favorable calibration and net benefit across clinically relevant thresholds. Our model-agnostic, open-source architecture enables independent validation and development. This work validates a pathway from automated measurement to clinical decision: reliable biomarkers drive both workload optimization today and patient risk stratification tomorrow, and the developed framework shows how foundation models can be operationalized within precision medicine systems.

💡 Research Summary

The authors present a comprehensive, end‑to‑end framework that transforms routine musculoskeletal (MSK) magnetic resonance imaging (MRI) into standardized quantitative biomarkers suitable for clinical decision support. Leveraging recent promptable foundation segmentation models—Segment Anything Model (SAM), its medical‑specific variant MedSAM, and SAM2—the study fine‑tunes these models on a heterogeneous collection of 12 MSK MRI datasets encompassing knee, hip, shoulder, lumbar spine, and thigh, totaling 913 scans acquired across multiple scanners (Siemens, GE), field strengths (1.5 T, 3 T), and protocols (2D spin‑echo, 3D DESS, quantitative T1ρ/T2, etc.).

Model selection deliberately restricts all three backbones to comparable parameter counts (≈80–90 M) to keep inference latency and GPU memory requirements realistic for clinical research environments. Each model is evaluated under a unified prompting regime: slice‑wise bounding‑box prompts generated automatically by an object‑detection network are fed to the mask decoder, ensuring consistent input across anatomies and eliminating the need for manual prompt engineering. For SAM2, the hierarchical memory module is disabled to avoid label occlusion in long volumetric sequences, thereby preserving deterministic behavior.

A YAML‑driven preprocessing pipeline normalizes intensities, resamples images to 1024 × 1024, stacks slices as pseudo‑RGB channels, and resizes masks to 256 × 256 with nearest‑neighbor interpolation and morphological closing. Data are split 70 %/15 %/15 % (train/validation/test) with subject‑level stratification by sex to maintain demographic balance. Segmentation performance is quantified using Dice similarity coefficient and Jaccard index, reported as median values per structure together with 5th‑percentile and minimum subject‑level scores to capture worst‑case behavior. Across all five anatomies, fine‑tuned models achieve Dice scores ranging from 0.86 to 0.92, with MedSAM showing the highest agreement for cartilage and bone.

The high‑quality segmentations serve as the measurement interface for a suite of quantitative biomarkers: cartilage thickness maps, bone height, muscle volume, and relaxation times (T1ρ/T2). Bland‑Altman plots and intraclass correlation coefficients (ICCs > 0.88) demonstrate that model‑derived measurements are statistically indistinguishable from expert manual annotations, establishing clinical fidelity.

Two downstream clinical applications illustrate the utility of these biomarkers. First, a three‑stage knee MRI triage cascade uses automatically derived cartilage and bone metrics to flag cases for radiologist review. Simulation shows a ~35 % reduction in verification workload while preserving >95 % sensitivity for pathology detection, thereby streamlining routine workflow without compromising safety. Second, a 48‑month longitudinal risk modeling pipeline leverages serial cartilage and meniscus thickness trajectories from the Osteoarthritis Initiative (OAI) cohort. Cox proportional hazards and random‑forest models predict individual risk of knee replacement and incident osteoarthritis. Calibration curves are favorable, and decision‑curve analysis reveals net clinical benefit across a range of risk thresholds, confirming that the biomarkers can support personalized prognostication.

The architecture is deliberately model‑agnostic and open‑source; code, pretrained weights, and detailed metadata are released on GitHub, enabling independent validation and extension to other MSK conditions such as spinal disc degeneration or rotator‑cuff tears. By decoupling the segmentation backbone from downstream analytics, the framework insulates clinical decision support from rapid model turnover, emphasizing stability at the point of care.

In summary, this work validates a practical pathway from foundation‑model‑based segmentation to reliable quantitative biomarkers, and demonstrates how those biomarkers can simultaneously deliver immediate operational efficiencies (triage workload reduction) and long‑term predictive value (risk stratification for osteoarthritis and joint replacement). The study provides a methodological blueprint for integrating advanced AI models into precision‑medicine pipelines, paving the way for scalable, vendor‑neutral deployment of quantitative MRI in routine musculoskeletal practice.