KWC-YOLO: An efficient YOLO architecture for lumbar spinal stenosis grading through dynamic convolution and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Lumbar spinal stenosis (LSS) poses a global health challenge, with MRI diagnosis facing interpretation variability and time constraints.
• Deep learning (DL) models for LSS show promise but are hindered by limited data, high computational needs, and difficulty capturing subtle features.

Purpose of the Study:

• To introduce KWC-YOLO, an efficient object detection framework for automated lumbar central canal stenosis (LCCS) grading using the Schizas criteria.
• To enhance DL model performance for LCCS detection by addressing data scarcity and computational demands.

Main Methods:

• Developed KWC-YOLO, an enhanced YOLOv11n architecture incorporating KernelWarehouse (KWConv) for dynamic convolution, FasterGATE activation for improved non-linearity, and a Slim-Neck structure for efficient feature fusion.
• Evaluated KWC-YOLO on a clinical lumbar spine MRI dataset.

Main Results:

• KWC-YOLO achieved a mean Average Precision of 86.7% (AP50) and 63.0% (AP95), outperforming the YOLOv11n baseline by 9.2% and 9.3% respectively.
• The model reduced computational load by 36.5% to 4.0 GFLOPs, demonstrating significant efficiency gains.
• Achieved superior performance in automated LCCS grading compared to existing methods.

Conclusions:

• KWC-YOLO sets a new benchmark for automated LCCS grading, offering high accuracy and computational efficiency.
• The framework has the potential to reduce radiologist workload, improve diagnostic accuracy, and streamline clinical decision-making for better patient outcomes.