SCS Detection using DeepscoreNet Architecture with Vertebra Localization via Yolov8

Abstract

Spinal Canal Stenosis (SCS), a condition marked by the pathological narrowing of the spinal canal, represents a significant and growing clinical challenge, particularly in aging populations. This narrowing frequently leads to compression of the spinal cord and nerve roots, manifesting as debilitating symptoms like chronic pain, sensory disturbances, and motor weakness, thereby severely impacting patient quality of life. Accurate diagnosis and, crucially, consistent severity grading using Magnetic Resonance Imaging (MRI) are paramount for guiding appropriate clinical management, ranging from conservative care to surgical intervention. However, manual MRI interpretation, while the current standard, is inherently time-consuming, requires substantial radiological expertise, and suffers from well-documented inter-observer variability, potentially leading to inconsistent treatment decisions. To mitigate these challenges and enhance diagnostic workflows, we propose a fully automated deep learning pipeline engineered for improved efficiency, objectivity, and reproducibility in SCS assessment. Our two-stage approach first employs the highly efficient YOLOv8 object detection model for rapid and precise localization of individual lumbar vertebrae (L1-S1) on standard sagittal T2-weighted MRI slices. This localization step provides crucial anatomical context for the subsequent, more detailed analysis. The second stage utilizes DeepScoreNet, a novel classification network we developed based on a VGG-19 backbone architecture. DeepScoreNet is specifically enhanced with integrated attention mechanisms (CBAM and SE blocks) to focus on diagnostically relevant image features. It performs fine-grained, multi- class severity grading (classifying stenosis into No Stenosis, Mild, Moderate, or Severe categories) independently for three critical anatomical regions often affected: Central Canal Stenosis (CCS), Left Foraminal Stenosis (LFS), and Right Foraminal Stenosis (RFS). We rigorously evaluated our pipeline’s performance using the publicly available Spider MRI Spine T2 dataset from the RSNA 2024 challenge [17]. The system demonstrated strong localization capability with a mean Average Precision (mAP@0.5) of 0.92 for vertebra detection. The DeepScoreNet classifier achieved a promising overall accuracy of 93% for severity grading, benefiting significantly from the use of focal loss to address the inherent class imbalance commonly found in medical datasets where severe cases are less frequent. To foster clinical trust and understanding, we incorporate Grad-CAM [16] visualizations (Figures 4-6) offering interpretability into the model’s predictive reasoning by highlighting salient image regions. Comprehensive ablation studies were conducted to systematically validate the positive contributions of both the focal loss function and the integrated attention modules to the final model performance. This paper provides a detailed account of the proposed system architecture, the underlying methodology including data handling and model training, extensive experimental results with performance metrics (Figures 2, 3, Table 4), and concludes with a thorough discussion of the findings, acknowledgments of current limitations, and potential avenues for future research and improvement.