CoordRefine-UNet: Enhancing MR Spine Image Segmentation with Coordinate Attention and Feature Refinement

In recent years, the need for spine segmentation (including multi-class segmentation of vertebrae and intervertebral discs) has become more urgent due to the rise in spine-related diseases. Accurate medical image segmentation can assist clinicians in diagnosing and treating spinal diseases more quickly and efficiently. However, traditional methods are time-consuming and labor-intensive, falling short of modern medical efficiency needs. To address this, we propose an efficient, automated spine MRI segmentation model based on UNet, enhanced with a Deep Coordinate Attention module (DCA) and Feature Refinement module (FRM), and using ResNet50 as the encoder for superior feature extraction. The DCA improves attention to critical areas by focusing on horizontal and vertical directions, while the FRM enhances detail capture in spine soft tissues and blurred edges, reducing mis-segmentation. With the ResNet50 encoder, the model handles complex anatomical structures more effectively. Our method improves IoU by 4.26% and DSC by 3.92% compared to UNet, and by 7.97% in IoU and 2.15% in DSC over ResNet50-UNet, demonstrating significant advancements in segmentation accuracy. These improvements showcase how the proposed modules enhance segmentation performance in spine MRI.