Single-Stage Deep Learning Pipeline for Multi-Label Ordinal Classification of Lumbar Spine Degenerative Disease from Multisequence MRI with Disc-Level Grad-CAM Explainability Analysis

Technical Summary: Single-Stage Deep Learning Pipeline for Multi-Label Ordinal Classification of Lumbar Spine Degenerative Disease

Problem Statement
Lumbar spine degenerative disease is a leading cause of global chronic disability, affecting approximately 266 million people. While MRI is the reference standard for diagnosis, expert interpretation is time-intensive and suffers from clinically significant inter-observer variability, particularly regarding neural foraminal narrowing. Existing deep learning solutions for lumbar classification are predominantly single-institution, address only one or two conditions, and lack disc-level explainability. Furthermore, no published study has simultaneously classified all five degenerative conditions across all five lumbar disc levels with ordinal severity grading while providing interpretable activation maps. This study addresses these gaps by developing a model capable of simultaneous ordinal classification for a complex 25-task problem (5 conditions × 5 disc levels) using multisequence MRI.

Methodology
The study utilized a retrospective, single-dataset multi-institutional design based on the RSNA 2024 LumbarDISC dataset, comprising 8,593 multisequence MRI series from 2,697 patients across eight institutions.

• Data Partitioning: Patient-level splitting was performed, resulting in a training set of 1,679 studies and an internal validation set of 296 studies.
• Input Processing: Multisequence MRI (sagittal T2, sagittal T1, and axial T2) was fused into a single 3-channel 512×512 JPEG input. The sagittal T2 used a mean projection of three central slices, the sagittal T1 used the middle slice, and the axial T2 used the middle slice by DICOM instance number.
• Model Architecture: A single-stage EfficientNet-B4 backbone (pretrained on ImageNet) was employed. The architecture outputs 75 logits, reshaped to [Batch, 25, 3], representing three-class predictions (Normal/Mild, Moderate, Severe) for each of the 25 condition-level pairs.
• Training Strategy: The model was trained using a distance-weighted cross-entropy loss with inverse-frequency class weighting to address grade imbalance. The loss function applied ordinal distance weights (1.5× for adjacent-grade errors, 2.0× for two-grade errors) to align with quadratic weighted kappa (QWK). Optimization was performed using AdamW with a learning rate of 3×10⁻⁵ over 32 epochs.
• Explainability: Gradient-weighted class activation mapping (Grad-CAM) was applied to the final convolutional block to generate heatmaps for the "Severe" class across all outputs. These were qualitatively reviewed on 10 validation studies with the highest Severe label counts.

Key Results

• Overall Performance: The model achieved a macro-averaged AUC of 0.61 for Severe disease detection and a macro-averaged QWK of 0.22 across all 25 outputs. According to Landis and Koch criteria, a QWK of 0.22 represents "slight-to-fair" agreement, which is below the moderate-to-substantial agreement typically seen between radiologists.
• Level-Specific Performance: Performance varied by disc level. The highest performance was observed at L3–L4 (mean QWK 0.32, mean AUC 0.65) and L4–L5 (mean QWK 0.24, mean AUC 0.70).
• Sensitivity vs. Specificity: At L4–L5, the model demonstrated high sensitivity for Severe spinal canal stenosis (0.965) but with a low specificity (0.569), resulting in a false positive rate of 43.1% among non-Severe cases.
• Condition-Specific Trends: Subarticular stenosis showed the highest mean kappa, while left neural foraminal narrowing showed the lowest (0.13), consistent with known radiologist variability in grading foraminal narrowing.
• Explainability: Grad-CAM analysis on 10 selected studies revealed anatomically plausible activation patterns. Heatmaps for Severe predictions localized progressively to lower lumbar levels (L3–S1) in multilevel disease, with appropriate lateralization for foraminal narrowing. No activation was observed outside the lumbar spine region.

Significance and Claims
The authors position this work as establishing a baseline for a complex 25-task ordinal classification problem rather than a clinically deployable tool.

• Baseline Establishment: The study demonstrates that a single-stage EfficientNet-B4 model can simultaneously classify five degenerative conditions across all five disc levels, a task more complex than previous single-condition or two-stage approaches.
• Explainability Contribution: The paper claims a novel contribution in providing disc-level Grad-CAM explainability for an MRI-based model operating across all five lumbar levels and five conditions simultaneously, a feature absent in prior classification literature.
• Clinical Limitations: The authors explicitly state that clinical deployment is not supported in the current form due to the low overall QWK (0.22) and the high false positive rate (43%) at high-sensitivity operating points. The results indicate the model is more effective at binary Severe detection than at fine-grained ordinal ranking.
• Future Directions: The study identifies key areas for improvement, including the use of two-stage pipelines with explicit disc-level patch extraction, prospective validation in low-resource settings (specifically at Patan Academy of Health Sciences), ensemble methods, and the adoption of dedicated ordinal regression losses (e.g., CORN or CORAL) to improve rank consistency.

In conclusion, while the model achieves clinically meaningful sensitivity for Severe disease at prevalent levels (L3–L4, L4–L5), the authors emphasize that further optimization and external validation are required before the system can be considered for practical clinical triage.