Automatic detection and voxel-wise mapping of lumbar spine Modic changes with deep learning

Area of Science:

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

Background:

• Modic changes (MCs) are common MRI findings in vertebrae, but standardized characterization, especially for mixed types, is lacking.
• Current methods struggle with definitive associations between MCs and low back pain, necessitating improved quantitative analysis.
• This study addresses the need for novel, interpretable AI tools for voxel-wise Modic change mapping.

Purpose of the Study:

• To develop and validate an interpretable deep learning model for detecting and mapping Modic changes in the lumbar spine.
• To create a voxel-wise 'Modic map' that standardizes the characterization of these vertebral anomalies.
• To assess the utility of the AI tool in improving radiologist agreement for MC assessment.

Main Methods:

• A retrospective study utilized 75 lumbar spine MRI exams. A two-stage deep convolutional neural network pipeline was employed.
• The first network segmented vertebral bodies, while the second, trained on radiologist-labeled MCs, performed MC segmentation.
• A rule-based algorithm categorized MC types, and an AI-assisted experiment evaluated radiologist performance with model predictions.

Main Results:

• The deep learning model achieved 85.7% accuracy in identifying MCs on an unseen test set, with a sensitivity of 0.71 and specificity of 0.95.
• A Cohen's kappa score of 0.63 indicated substantial agreement for the model's detection of MCs.
• AI-assisted assessment significantly improved inter-rater reliability between junior and senior radiologists (kappa from 0.52 to 0.58).

Conclusions:

• The developed deep learning tool shows significant agreement with human radiologists in identifying Modic changes.
• This AI-based approach offers a promising method for enhancing inter-rater reliability in spinal MRI assessments.
• The interpretable voxel-wise mapping may aid in more standardized and consistent diagnosis of vertebral anomalies.