DERNet — Spinal Lesion Analysis

Problem

Spinal lesion detection from X-ray images is a high-volume clinical task with significant inter-reader variability. Automated, consistent detection systems can reduce radiologist workload while flagging subtle pathologies that may be missed under time pressure.

Approach

DERNet (Deep Ensemble Residual Network) addresses spinal lesion detection on the VinDr-SpineXR radiograph dataset. The architecture uses a residual backbone enhanced with spatial attention modules and is trained as an ensemble of independently-initialized models, enabling prediction-time uncertainty estimation through disagreement.

Key components:

• ResNet-based backbone with squeeze-and-excitation attention
• Multi-scale feature pyramid for detecting lesions of varying sizes
• Ensemble diversity training to maximize coverage of the prediction space
• Grad-CAM++ heatmaps for localization interpretability

Results

Evaluated on the VinDr-SpineXR dataset:

• mAP@0.5: 0.74
• Strong performance on common lesion classes (fracture, disc herniation)
• Uncertainty maps correlate with rare or ambiguous cases

Reproducibility

Full pipeline available in repository: data preprocessing, training, ensemble inference, and Grad-CAM visualization scripts per the CIMILab Reproducibility Standard.