Fed-SpineNet: Federated Transformer Framework for Multi-Center Privacy-Preserving Lumbar MRI Analysis

Lumbar degenerative conditions are a major cause of chronic low back pain, and MRI is the primary tool for assessing their severity. Grading across multiple lumbar levels is slow and inconsistent between readers. Deep learning models trained at single sites or on pooled data struggle to generalize across hospitals with different scanners and protocols, and sharing raw scans between institutions is rarely feasible. Fed-SpineNet is a federated learning framework that classifies multi-level lumbar degeneration severity while keeping patient data local. It fuses sagittal T1, sagittal T2/STIR, and axial T2 sequences using cross-view attention, and incorporates center-adaptive normalization, ordinal severity loss, and attention consistency regularization. On the LumbarDISC dataset of 2,697 patients and 8,593 MRI series from eight centers, it achieves 89.4% balanced accuracy, 88.1% macro F1-score and 0.968 AUC, surpassing all local, centralized and federated baselines. These findings suggest that federated training of task specific components can match centralized performance in spinal pathology grading.