Landmarks Detection and Parameters Calculation of Full Spine X-ray Image Based on U-Net

Scoliosis is a common spinal deformity that significantly impacts patients' physical and mental health. Comprehensive assessment of scoliosis is crucial for developing effective treatment plans. Therefore, we collected a dataset consisting of 853 full-spine X-ray images and developed an artificial intelligence-based solution for spine-pelvis landmark detection in full spine X-ray images, utilizing the U-Net. We innovatively transformed the detection task into a segmentation task of circles centered on the landmarks with a diameter of 60 pixels, and the centroid of the segmented circles are the predicted landmarks. The analysis of landmark detection results showed a minimum average distance error of 1.85mm. Subsequently, based on the detected landmarks, we calculated a series of spine-pelvis parameters commonly used in clinical scoliosis examinations. The symmetric absolute percentage error analysis yielded a minimum parameter error of 12.28%. Linear regression analysis between the predicted parameters and their true values showed a significant linear relationship between the predicted and true values, with the clavicle angle parameter achieving the highest correlation coefficient of 0.993.