Automatic spinal Cobb angle measurements from X-ray images using a novel vertebra centroid radiation landmark detection network

Adolescent idiopathic scoliosis is a common spinal disease that can be evaluated by calculating the Cobb angle. Manual Cobb angle calculation is subjective and time-consuming. However, automated Cobb angle prediction methods have limitations, such as prediction errors for vertebrae and inaccurate landmark detection. Accordingly, we propose an improved two-stage approach that first applies YOLOv7 to screen and segment each of the target vertebrae for landmark prediction. Subsequently, to improve landmark detection accuracy, the approach applies our developed multibranch network, denoted as Vertebra Centroid Radiation Landmark Detection Network (VCRLD-Net), which includes a Long-Range Squeeze Omnidimensional Attention (LSOA) module, for individual vertebra analysis. Experiments revealed that VCRLD-Net achieved a mean distance error of 8.82, a circular mean absolute error for angles of 2.38°, and a symmetric mean absolute percentage error of 5.01%, outperforming previously proposed methods.