RASpine: Regional Attention Lateral Spinal Segmentation based on Anatomical Prior Knowledge.

In the clinical diagnosis and treatment of spinal disorders, segmenting the spine from X-ray images provides clear visualization of the spinal structure and morphology. However, while existing spine segmentation methods perform well on anteroposterior X-ray images, their performance is poor on lateral X-rays. This is mainly due to the low contrast and severe occlusion of the thoracic vertebrae on lateral X-rays, resulting in overlapping vertebrae in segmentation results. To address this issue, this paper proposes a segmentation network called Region Attention and Spine Prior-based Network (RASpine). By utilizing the anatomical prior knowledge of non-overlapping regions between different vertebrae, an overlap detector is designed to identify overlapping parts of different vertebrae in the segmentation results. Moreover, a loss function is designed to penalize the overlapping regions, thereby avoiding overlapping segmentation results for the vertebrae. Finally, region attention is employed to enhance the segmentation accuracy in challenging regions. The proposed RASpine is trained, validated, and tested on a clinical dataset. Experimental results demonstrate that compared to existing mainstream medical image segmentation algorithms, RASpine effectively addresses the overlapping parts in lateral X-ray spine segmentation results and achieves more satisfactory performance in multiple evaluation metrics.