Spinal Stenosis Grading in Magnetic Resonance Imaging Using Deep Convolutional Neural Networks

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Spinal Diagnostics

Background:

• Spinal stenosis grading from MRI is crucial but often subjective and time-consuming.
• Inter-observer variability among experts can lead to diagnostic discrepancies.
• Current methods require manual review of numerous image slices.

Purpose of the Study:

• To assess the feasibility of a computer-assisted system for spinal stenosis grading.
• To compare diagnostic agreement between human experts and deep convolutional neural network (CNN) classifiers.
• To evaluate the reliability of AI in spinal stenosis assessment.

Main Methods:

• Retrospective analysis of 542 L4-L5 axial MRI scans.
• Two experts independently graded spinal stenosis.
• Two CNN classifiers (Faster R-CNN for detection, VGG for classification) were trained and validated using expert labels via 10-fold cross-validation.

Main Results:

• Expert-expert agreement showed 77.5% accuracy and 75% F1 score.
• Expert-model agreement reached 83% accuracy and 75.4% F1 score (Expert 1), and 77.9% accuracy and 74.9% F1 score (Expert 2).
• The differences between expert and AI model grading were not statistically significant.

Conclusions:

• Automatic spinal stenosis grading using deep learning is feasible.
• AI-based systems demonstrate diagnostic agreement comparable to human experts.
• This technology holds promise for improving the efficiency and consistency of spinal stenosis diagnosis.