Automated deep learning pipeline for measuring lumbar thecal sac AP diameter on mid-sagittal MR images

To develop and validate an automated, disc-level deep learning pipeline for quantitative measurement of anteroposterior (AP) thecal sac diameter on mid-sagittal lumbar T2-weighted MRI.

In this retrospective study, 511 mid-sagittal lumbar T2 MRI examinations were included after screening 758 cases and applying predefined exclusions. The workflow combined YOLOv8 oriented bounding boxes (OBB) for disc-level localization and orientation estimation, homography-based ROI warping, Attention U-Net segmentation, and skeleton-based AP diameter computation in millimeters using DICOM pixel spacing. Validation was performed on internal (50) and external (50; RSNA 2024 lumbar dataset) cohorts with two radiologists providing the reference standard.

Inter-reader agreement was excellent (ICC (2, 1) = 0.967; 711 paired measurements). Against the reader-mean reference, the pipeline achieved an overall MAE of 0.994 mm (711 disc-level measurements). Internal validation showed MAE 0.909 mm (357 measurements) and external validation MAE 1.079 mm (354 measurements). Severity-wise MAE remained ~ 1 mm (mild 0.930 mm; moderate 1.234 mm; severe 1.038 mm). Automatic disc-level labeling was performed, and OBB-derived orientation significantly improved AP measurement-line validity versus axis-aligned detection (acceptable lines 99.02% vs. 77.64%).

An orientation-aware YOLOv8-OBB + Attention U-Net pipeline enables automated, disc-level AP thecal sac diameter quantification on mid-sagittal lumbar MRI with ~ 1 mm error relative to expert reference, supporting standardized morphometric reporting and measurement-driven assessment of lumbar stenosis.