A Novel Lumbar Foraminal Stenosis Detection and Classification Framework Using 3D Sagittal MRI Dataset

Abstract

Lumbar foraminal stenosis (LFS) is a spinal condition commonly resulting from degenerative changes in the foraminal region of the lumbar spine, leading to nerve compression and symptoms such as leg pain and muscular weakness. Conventional diagnosis of LFS through manual MRI assessment by radiologists and neurologists is both timeintensive and prone to inter-observer inconsistencies. To enhance diagnostic efficiency and consistency, we propose an automatic AI-driven framework that identifies optimal MRI slices containing visible LFS indicators and performs comprehensive analysis on 3D sagittal MRI DICOM volumes. The framework comprises three key components: (1) Slice selection via a custom 3D convolutional neural network (CNN), (2) region of interest (ROI) detection via the adopted YOLOv12, and (3) Foraminal stenosis grading using the proposed hybrid classification module using the late fusion strategy by combining the custom CNN with the DeiT-based vision transformer decisions. The system demonstrates promising performance, achieving an accuracy of 86% in slice selection, 97.2% in ROI detection, and 65% in LFS grading across the lumbar spine regions. These results suggest that the proposed pipeline offers a practical solution for automated, end-to-end assessment of lumbar foraminal stenosis using sagittal MRI data. © 2025 IEEE.