Segmentation and Measurements of Clinically Relevant Anatomy of the Lumbar Spine from the MRI Sagittal View

Le résumé fourni par la source

Accurate assessment of the lumbar spine requires precise segmentation, reliable vertebral localization, and reproducible morphometric measurements. Current workflows depend heavily on manual annotation, which is slow and difficult to scale. Deep learning models offer promise but often overfit small datasets, while regressionbased methods typically predict only a limited set of accurate measurements. We present an end-to-end pipeline that automates segmentation, localization, and measurement of all clinically relevant lumbar metrics from sagittal MRI scans. The framework combines MU-Net+ for multiclass segmentation, YOLOv8/YOLOv5m for rapid vertebral detection, and a deterministic Automated Anatomical Spinal Measurement Module (AASMM) that extracts 24 angular and distance-based measurements directly from segmented and localized vertebrae. Our evaluation shows that MU-Net+ achieved a pixel accuracy of 99.21%, mean Intersection over Union (IoU) of 87.40%, and Dice Similarity Coefficient (DSC) of 98.43%. YOLOv8 reached 99.6% precision and 99.4% mAP for localization. The AASMM module demonstrated an overall mean Pearson Correlation Coefficient (R) of 0.779% and a mean absolute error (MAE) of 1.96 mm across 24 measurements, with strong agreement to manual readings, with key metrics R = 0.92 for the Lumbar Lordosis Angle and R = 0.87 for anterior vertebral height between L3 and L4. These results highlight a fast, accurate, and reproducible pipeline for lumbar spine analysis. By minimizing manual input and enhancing scalability, the proposed framework represents a practical step toward routine, automated spine assessment in clinical radiology.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.