Anatomy-guided context-aware deep learning for lumbar degenerative disease grading and burden-aware risk assessment on MRI.

Automated lumbar spine MRI analysis has made notable progress in recent years, yet existing methods still suffer from anatomical ambiguity, insufficient use of explicit structural priors, and limited modeling of inter-level degenerative dependency, which restrict their reliability for clinically meaningful burden assessment. To address these limitations, we propose an anatomy-guided, multi-sequence, multi-level deep learning framework for lumbar degenerative disease grading and patient-level risk assessment. First, an anatomical parsing module is pre-trained to segment vertebral bodies, intervertebral discs, and the spinal canal, providing stable level localization and structural priors for downstream analysis. Then, the localized multi-sequence MRI patches and quantitative anatomical biomarkers are jointly encoded and further modeled by a lightweight Transformer to capture contextual dependency across lumbar levels, enabling both segment-level ordinal grading and Clinically Significant Degeneration Score (CSDS)-based patient-level burden assessment. Extensive experiments on public datasets demonstrate that the proposed framework achieves superior performance over representative baselines, reaching a Macro F1-score of 0.783 ± 0.010, a Cohen's Kappa of 0.765 ± 0.012, a weighted log loss of 0.463 ± 0.018, and a patient-level AUC of 0.891 ± 0.009, while supplementary evaluations further verify the robustness of the anatomical parsing stage. The ablation results show that anatomy-guided cropping, quantitative biomarker fusion, spine-context modeling, and consistency regularization each contribute complementary gains, confirming that the effectiveness of the framework arises from their coordinated interaction rather than from any single component alone. Overall, these findings indicate that explicitly integrating anatomical priors, structured biomarkers, and multilevel context can substantially improve both diagnostic accuracy and interpretability, while also supporting more standardized, transparent, and human-centered lumbar MRI assessment in radiological practice.