AOI-LSCS: Angle-Oriented ROI for Accurate Detection and Classification of Lumbar Spinal Canal Stenosis Intervertebral Disc Levels

This study presents a methodology for the detection and classification of intervertebral disc levels in lumbar spinal canal stenosis (LSCS) using sagittal T2-weighted magnetic resonance images (MRI), which provide high contrast for visualizing spinal structures. The dataset, curated by the Radiological Society of North America (RSNA), includes annotations marking the disease center. A robust angle-oriented region of interest (ROI) extraction technique is introduced, leveraging the disease center and incorporating precise angle correction to align the region of interest with the spinal curve. Bounding boxes are derived from the annotations and adapted to the DOTA format to ensure compatibility with OpenMMLab’s MMRotate framework. Utilizing ResNet50 as backbone with Feature Pyramid Network (FPN) architecture, the approach is fine-tuned to accurately classify intervertebral disc levels, including L1/L2 through L5/S1. The proposed pipeline achieves 76.4% accuracy, demonstrating its potential for reliable and precise identification of the disc level in clinical applications.