Deep learning-based fully automated vertebral endplates irregularity prediction using lumbar magnetic resonance imaging

Abstract

Introduction: Currently, the clinical analysis of lumbar MRI heavily relies on the manual and subjective assessment process. It is inefficient and inconsistent and unable to predict the longitudinal pathology progression. Therefore, we aimed to establish a deep learning-based system for automated analysis of lumber MRI. Methods: The dataset was developed on 1152 volunteers (40.17% male) from the southern Chinese population, who had a mean age of 41.43 years, and the main age-group was 40 to 50 years (42.82%). For each volunteer, two MRI scans (baseline and follow-up) were collected. Three different pathologies of vertebral endplates irregularity, including Schmorl’s node, high intensity zones (HIZs), and marrow change, were assessed by two spine specialists with over ten years of clinical experience. Our deep learning-based system integrated the published MRI-SegFlow to segment spinal tissues and a convolutional neural network to predict follow-up pathologies based on the baseline MRI and segmentation results. The 5-fold cross-validation was conducted for the quantitative validation of our system. Results: Validation results showed that our system achieved remarkable performance on the pathology prediction of Schmorl’s node (mean accuracy: 89.46 ± 3.71%), HIZ (mean accuracy: 91.75 ± 2.48%), and marrow change (mean accuracy: 87.51 ± 2.23%). Discussion and Conclusion: A deep learning-based system for fully automated lumbar MRI analysis is implemented and tested. The validation results show that the system can achieve remarkable performance on the prediction of multiple vertebral endplates irregularity pathologies. Our system has significant potential for clinical implementation.