A compact teacher-student knowledge-distillation framework for lumbar MRI segmentation

Accurate segmentation of lumbar magnetic resonance imaging (MRI) is important for quantitative assessment of degenerative spine disease, but it remains challenging because spinal structures show low contrast, variable morphology, and severe pixel-class imbalance. This study evaluates a compact teacher-student knowledge-distillation framework for multi-class lumbar MRI segmentation. The student network integrates a Res2Net encoder-decoder, a lightweight transformer bottleneck, and deep supervision; a manifold-constrained hyper-skip (MHC) module is evaluated as an optional constrained skip-fusion variant rather than assumed to be the sole source of improvement. A UNet++ teacher transfers soft probability, class-focused, and boundary-weighted information to guide boundary-sensitive anatomical regions. Under five-fold case-level GroupKFold validation on the primary public axial lumbar MRI dataset, the MHC full-KD variant achieved Dice 0.9141 ± 0.0021, foreground Dice 0.8934 ± 0.0026, IoU 0.8519 ± 0.0033, and HD95 3.1451 ± 0.2101. A no-MHC student trained with the same full distillation objective achieved essentially the same Dice (0.9141 ± 0.0021) with lower HD95 and lower measured latency, indicating that the full distillation objective, rather than the MHC block alone, explained most of the performance recovery toward the teacher. Replication on the 2026 Medical Lumbar Spine 3D Axial MRI Dataset achieved Dice 0.8960 ± 0.0032 and IoU 0.8224 ± 0.0047. A fixed label-harmonized train-on-primary/test-on-2026 transfer stress test achieved macro Dice 0.7002 ± 0.0181 and foreground Dice 0.6280 ± 0.0225. The results suggest that compact class-aware distillation can support reproducible lumbar MRI segmentation for future computer-aided measurement pipelines, while prospective and strict cross-institutional validation remain necessary.