Deep learning models for lumbar spinal stenosis on MRI: model comparison and clinical benchmarking

Area of Science:

• Radiology
• Artificial Intelligence
• Spine Surgery

Background:

• Lumbar spinal stenosis diagnosis relies on imaging interpretation.
• Automated classification using deep learning offers potential for improved efficiency and accuracy.

Purpose of the Study:

• To compare deep learning models (CNN and transformer-based) for automated lumbar spinal stenosis classification on MRI.
• To benchmark their performance against radiologists and orthopedists.

Main Methods:

• Retrospective analysis of 564 lumbar spine MRI studies.
• Development of CNN and transformer models for stenosis classification.
• Performance evaluation against 8 clinicians using detection recall, sensitivity, specificity, and interrater agreement (Gwet κ).

Main Results:

• Both deep learning models achieved high recall (>94%) for all regions of interest.
• CNN and transformer models demonstrated high agreement (κ=0.90-0.99) with expert consensus for stenosis classification, comparable or superior to clinicians.
• Performance was consistent across internal and external test sets.

Conclusions:

• Deep learning models, particularly CNNs, show comparable or superior performance to clinicians in classifying lumbar spinal stenosis from MRI.
• These AI tools have the potential to assist in clinical report generation, surgical planning, and medical education.