Residual multi-scale attention and deformable learning for lumbar spine MRI segmentation

Segmentation of lumbar spine structures (vertebrae, intervertebral discs, and spinal canal) from magnetic resonance images remains a difficult task due to low tissue contrast, anatomical variability across patients, and the high computational cost of modern deep learning models. In the present study, we propose implementing Residual Multi-Scale Attention with UNet+ + (RMSA + UNet + +). The recommended segmentation model focuses on UNet+ + and includes three major architecture-related improvements: Residual Learning, a feature that helps facilitate the reliable propagation process of conventional gradients; Multi-Scale Feature Extraction, a technique that allows for context-aware boundary detection at multiple spatial scales; and deformation-based attention, which dynamically localises the unbalanced anatomical boundary lines. A systematic preprocessing pipeline incorporating normalisation, anisotropic diffusion, CLAHE, and data augmentation improves Dice from 0.812 on raw images to 0.912 after full preprocessing, a gain of 10.0 percentage points. Evaluated on the SPIDER multi-centre LS-MRI benchmark across 218 patients from four hospital sites, RMSA + UNet+ + achieves a Dice of 0.947 ± 0.002, an IoU of 0.924 ± 0.003, and an ASD of 0.87 ± 0.02 mm for vertebrae segmentation, outperforming UNet+ + , Dense UNet+ + , Attention UNet+ + , and Swin UNet+ + across all three anatomical targets. Performance gains over all 4 baselines are statistically significant at p < 0.05 on per-subject Dice scores. The lightweight L = 1 configuration achieves 6.3 M parameters and 12.7 G FLOPs, against 23.5 M parameters and 67.3 G FLOPs for Swin UNet+ + , while maintaining Dice within 1% of the full model. Ablation studies confirm independent and synergistic contributions of each network component, and convergence by epoch 50, with aligned training and validation curves, indicates stable learning without overfitting. These results found RMSA + UNet+ + as a competitive method for LS-MRI segmentation on the SPIDER benchmark, with a lightweight variant suitable for resource-constrained deployment.