OpenMSK: A modular open-source framework for comprehensive analysis of imaging biomarkers in knee osteoarthritis

To develop and validate a fully automated, extendable, open-source AI-driven pipeline capable of knee magnetic resonance imaging (MRI) segmentation and quantitative extraction of MRI and positron emission tomography (PET) imaging biomarkers.

For model training, open-source knee MRI data from 347 knees acquired across 7 scanners using multiple gradient-echo sequences were used. An independent test set of 20 knees from 3T PET-MRI examinations were used for evaluation in this prospective reliability and agreement study. The pipeline segmented femoral, tibial, and patellar bones and cartilage, and medial and lateral menisci using three different deep learning models. Quantitative outputs included cartilage thickness and T₂ relaxation times, meniscus volume and T₂∗ relaxation times, bone shape scores, and static and kinetic subchondral bone PET measures. Manual segmentations from two annotators served as reference standards. Segmentation performance was assessed using Dice similarity coefficients (DSC), average symmetric surface distances, and volume differences. Quantitative outcome agreement was evaluated using intraclass correlation coefficients (ICC), Pearson correlations, and root mean square error.

Automated segmentations demonstrated high spatial agreement with manual annotations, with DSCs ranging from 0.84 to 0.98 across all tissues. Quantitative PET-MRI biomarkers generally demonstrated high agreement, although confidence intervals were wide for several biomarkers.

OpenMSK enables multi-tissue knee segmentation and generation of quantitative MRI and PET biomarkers, with performance assessed against two independent annotators and across different U-Net-based segmentation architectures. As a freely available, modular platform, OpenMSK lowers technical barriers to whole-joint quantitative imaging and provides a standardized foundation for OA research and clinical translation.