Automated Spinal Curvature Assessment from X-Ray Images Using Landmarks Estimation Network via Rotation Proposals

Adolescent idiopathic scoliosis (AIS) is one of the most common type of scoliosis. In current clinical settings, the severity of scoliosis is assessed by evaluating the contralateral blending angle of the spinal cord. Cobb angle is one of the most widely accepted standards for angle measurement. However, the manual measurement of Cobb angle is time consuming and unreliable. In this article, we propose a novel two-stage method that can automatically estimate Cobb angle from vertebrate landmarks. The proposed method uses rotation vertebrate region proposals to increase the accuracy of vertebrate localization in curved spinal region. Our model uses a backbone of ResNet50 combined with FPN for multiscale region proposal extraction. The rotation proposals are co-registered and fed into stage-two fully convoluted network (FCN) for vertebrate landmarks detection. The performance of proposed method is more robust than traditional landmarks segmentation networks for datasets with large variance, with a SMAPE score of 25.4784.