Related Experiment Video
Updated: Aug 30, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Structured MOAKS-based MRI features for predicting medial joint space narrowing progression: a comparative
Minggui Bao1, Ahmad Alkhatatbeh1, Jiankun Xu2
1Department of Orthopedics Surgery, The First Affiliated Hospital of Shantou University Medical College, Shantou, China.
Background:
Knee osteoarthritis (OA) is a major cause of pain and disability, yet baseline radiographic severity grading provides only modest prediction of structural progression. We evaluated whether structured MRI-derived MOAKS features improve prediction of full-grade medial joint space narrowing (JSN) progression, and whether deep multimodal fusion adds value beyond structured feature models.
Methods:
We performed a patient-independent internal-validation study using the Osteoarthritis Initiative. Full-grade medial joint-space-narrowing (JSN) progression was defined as an increase of at least one grade between baseline and fixed 48-month central radiographic readings from the same longitudinal reading project. The final analytic cohort included 2,899 knees from 2,392 participants, partitioned at the participant level into training, validation, and held-out test sets of 2,050, 421, and 428 knees. We compared a radiographic baseline model (RBM), raw-MOAKS model (RMO), structured MRI/radiographic model (SMR), and radiograph-plus-MOAKS ResNet50 fusion model (XMF). AUPRC was primary; uncertainty was estimated with 2,000 patient-cluster bootstrap replicates.
Results:
Progression prevalence was 10.0%. On the held-out test set (40 progressor knees), AUPRC/AUROC were 0.220 (95% CI 0.142-0.331)/0.748 (0.663-0.827) for RBM, 0.412 (0.260-0.576)/0.805 (0.726-0.878) for RMO, 0.345 (0.219-0.507)/0.801 (0.725-0.871) for SMR, and 0.301 (0.191-0.464)/0.768 (0.679-0.852) for XMF. Compared with RBM, AUPRC increased by 0.192 (0.087-0.313) for RMO and 0.125 (0.042-0.233) for SMR; the corresponding AUROC differences were 0.057 (-0.012 to 0.123) and 0.053 (0.004-0.105). XMF did not improve on SMR in AUPRC or AUROC (differences -0.044 [-0.135 to 0.060] and -0.033 [-0.092 to 0.018], respectively). A stronger clinical-radiographic comparator achieved AUPRC/AUROC 0.251/0.759, increasing to 0.355/0.813 after adding MOAKS.
Conclusion:
Baseline MOAKS features provided prognostic information beyond radiographic or clinical-radiographic severity for fixed 48-month medial JSN progression. In this selected MOAKS-scored cohort, adding a baseline-radiograph ResNet50 encoder did not improve discrimination over the structured models. The finite number of test events, selected MOAKS availability, calibration requirements, and lack of external validation preclude claims of clinical readiness.
