Objective:

To propose an anatomy-guided, multi-sequence, multi-level deep learning framework for grading lumbar degenerative diseases and assessing patient-level risk via MRI.

Approach:

• Anatomical Parsing Module: Pre-trained to segment vertebral bodies, intervertebral discs, and the spinal canal for stable level localization and structural priors.
• Multi-Sequence MRI Patches: Localized patches and quantitative anatomical biomarkers are jointly encoded and modeled by a lightweight Transformer to capture contextual dependency across lumbar levels.
• Grading and Assessment: Enables segment-level ordinal grading and Clinically Significant Degeneration Score (CSDS)-based patient-level burden assessment.

Key Findings:

• TheproposedframeworkachievedaMacroF1-scoreof0.783±0.010,aCohen'sKappaof0.765±0.012,aweightedloglossof0.463±0.018,andapatient-levelAUCof0.891±0.009.Additionally,anatomy-guidedcropping,quantitativebiomarkerfusion,spine-contextmodeling,andconsistencyregularizationeachcontributedtoperformancegains.

Interpretation:

The integration of anatomical priors, structured biomarkers, and multilevel context is associated with improvements in diagnostic accuracy and interpretability in lumbar MRI assessment.

Limitations:

• Existingmethodsstillfacechallengesrelatedtoanatomicalambiguityandinter-subjectstructuralvariation.Furthermore,manystudiesfocusonisolatedsegment-levelassessmentswithoututilizingexplicitanatomicalpriors.

Conclusion:

The proposed framework demonstrates enhanced diagnostic performance and interpretability, contributing to more standardized lumbar MRI assessments.