Abstract
Spinal Canal Stenosis is a prevalent condition that occurs as the spaces within the spinal canal gradually narrow over time due to degenerative changes in the ligaments, joints, and bones. This can lead to chronic pain, stiffness, and limited movement, significantly affecting a person's daily activities and overall well-being. Traditional methods for classifying Spinal Canal Stenosis can be slow and prone to errors. To address these challenges, this study proposes a deep learning-based approach utilizing Convolutional Neural Networks (CNNs) to classify Spinal Canal Stenosis into three severity levels: Normal/Mild, Moderate, and Severe. Using the RSNA 2024 Lumbar Spine Degenerative Classification Challenge dataset, which consists of a set of different types of Lumbar MRIs, axial T2, Sagittal T1, and Sagittal T2/STIR. For this study, we decided to use axial T2-weighted MRI scans. We implemented a series of preprocessing techniques including DICOM-to-PNG conversion, K-Means clustering for slice selection, and data augmentation to address class imbalance. The CNN model, designed with five convolutional blocks and enhanced with batch normalization, dropout regularization, and an early stopping mechanism, achieved an overall classification of 89% with a high recall of 97% for the Severe category. These findings indicate that deep learning models can substantially enhance the accuracy and efficiency of diagnosing lumbar spinal canal stenosis and other degenerative spine conditions, offering valuable support to radiologists. Also, we will explore the integration of Vision Transformers (ViTs) and multimodal learning techniques to further enhance the model’s performance and clinical applicability.