Multi-class segmentation and deep radiomics analysis of multi-center lumbar spine MRI using wavelet fusion and explainable machine learning

Accurate, reproducible assessment of lumbar spine pathologies is essential for clinical management but remains challenging due to inter-observer variability. We aimed to develop a comprehensive pipeline integrating multi-class segmentation and radiomics classification to enhance diagnostic precision for degenerative disc disease (DDD) and spinal canal stenosis. A retrospective multi-center cohort of 2365 patients provided T1-and T2-weighted sagittal lumbar MRI images. MRI sequences underwent wavelet-based fusion, enhancing structural clarity and contrast. Five state-of-the-art segmentation models (nnU-Net, TransUNet, Swin-UNet, Attention U-Net, and MedSAM) were evaluated for delineating lumbar vertebrae, intervertebral discs, and the spinal canal. Radiomic features (n = 215) extracted from segmented structures underwent correlation filtering and feature selection using Mutual Information (MI), Recursive Feature Elimination (RFE), and ANOVA. Classification tasks included Pfirrmann grading, stenosis severity, and patient-level DDD diagnosis. Performance was validated using an independent external cohort (n = 753) and evaluated via Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), precision, recall, volumetric similarity, accuracy, and AUC metrics. Segmentation accuracy peaked with Swin-UNet using fused MRI (mean DSC = 94.5 %; HD = 4.1 mm), outperforming all comparator models. Radiomics classification with MI-selected features and Ensemble Voting achieved high accuracy (Pfirrmann grading = 91.2 %, stenosis severity = 89.8 %, patient-level diagnosis = 91.0 %) and robust AUC (>92 % for internal tests, ∼90 % externally). SHAP analysis revealed critical radiomic features associated with clinical severity across all classifications, underscoring strong interpretability. Our integrated approach significantly enhances lumbar MRI analysis, offering superior segmentation and classification performance. This pipeline has excellent potential for clinical adoption, improving diagnostic accuracy, consistency, and decision-making for lumbar spine disorders.