Automated detection of lumbar disc herniation at L4–L5 and L5–S1 levels on sagittal MRI using a YOLO-based deep learning model

To conduct a preliminary single-center feasibility study of a YOLO-based deep-learning model for automated detection of lumbar disc herniation at the L4–L5 and L5–S1 levels on sagittal MRI. In this retrospective study, 372 anonymized sagittal T2-weighted lumbar MRI slices from adult patients evaluated for low back pain at a single tertiary center were reviewed. Intervertebral discs at the L4–L5 and L5–S1 levels were labeled with bounding boxes into four classes (L4–L5 herniated, L5–S1 herniated, L4–L5 non-herniated, L5–S1 non-herniated) by two clinicians (a radiologist and a PM&R specialist, each with 6 years of post-residency experience), with consensus adjudication of disagreements, using the combined North American Spine Society / ASSR / ASNR nomenclature as reference. The dataset was split into 337 training, 25 validation, and 10 test images. Rotation-based augmentation was applied. A YOLO object-detection architecture was trained and evaluated using precision, recall, F1, and mean average precision (mAP@0.5 and mAP@0.5–0.95). 95% bootstrap confidence intervals were estimated for aggregate metrics. The model achieved an overall precision of 0.738, recall of 0.698, mAP@0.5 of 0.744, and mAP@0.5–0.95 of 0.454, with a best overall F1 of 0.69 at a confidence threshold of approximately 0.30. Herniated classes outperformed non-herniated classes, with the highest recall observed for L5–S1 herniated discs (0.963). Bootstrap confidence intervals were wide, consistent with the small test set. This preliminary feasibility study suggests that YOLO-based level-specific detection of lumbar disc herniation at L4–L5 and L5–S1 on sagittal MRI is technically feasible. Given the very small test set, single-center design, absence of external validation, and lack of radiologist benchmarking, these results are hypothesis-generating and not yet sufficient to support any clinical or autonomous-diagnostic use. Larger multi-center validation and prospective comparison with radiologists are required.