A team of researchers from the UCSF Center for Intelligent Imaging (UCSF ci2) developed a modular system that converts routine musculoskeletal MRI into standardized quantitative biomarkers for clinical decision support.
The researchers, including UCSF ci2's Gabrielle Hoyer, Michelle Tong, Rupsa Bhattacharjee, PhD, Valentina Pedoia, PhD, and Sharmila Majumdar, PhD present the findings in "Clinical utility of foundation models in musculoskeletal MRI for biomarker fidelity and predictive outcomes," published in npj Digital Medicine.
The study addresses a persistent gap in musculoskeletal radiology. Generating quantitative biomarkers from MRI typically requires trained experts to manually delineate multiple structures across hundreds of slices, a time-intensive process that limits scalability in clinical practice. The team built a modular system that fine-tunes promptable foundation segmentation models, specifically SAM, SAM2, and MedSAM, across heterogeneous musculoskeletal datasets spanning five anatomies: knee, hip, shoulder, lumbar spine, and thigh.
The results were notable. Fine-tuning raised segmentation accuracy substantially, with 83% of fine-tuned scores reaching a Dice coefficient of 0.90 or above. Biomarker agreement with expert measurements was high across cartilage thickness, bone volume, disc height, muscle volume, and relaxometry, with intraclass correlation coefficients reaching as high as 0.99. For cartilage thickness specifically, Bland-Altman limits of agreement were within plus or minus 0.18 mm, comparable to published test-retest precision for MRI-based cartilage morphometry.
The team then applied these validated biomarkers to two clinical tasks. The first was an automated three-stage triage cascade for knee MRI that reduced radiologist verification workload to as low as 1.7 hours per 1,000 scans while maintaining sensitivity for clinically significant pathology. The second was a set of 48-month landmark models forecasting knee replacement and incident osteoarthritis, with the knee replacement model reaching an AUC of 0.76 and favorable calibration across clinically relevant thresholds.
The team notes that their model-agnostic, open-source architecture enables new segmentation backbones to be integrated without affecting biomarker extraction or downstream decision layers, a design choice intended to support sustainable clinical deployment over time.
Visit the UCSF ci2 blog to read more about research and news.
