Automated Cobb Angle Measurement Using MVIE-Net Combined with Vertebral Segmentation and Landmarks Detection

The Cobb angle is the most widely used measurement to quantify the magnitude of scoliosis. Accurate automated measurement of the Cobb angle can improve the efficiency of scoliosis diagnosis. The existing direct estimation of Cobb angle cannot extract structural information of the spine and lacks interpretability. Curvature-based Cobb angle estimation rely on vertebral feature information tend to focus on a single landmark or segmentation information and cannot provide robust vertebral feature information for post-processing of curvature calculations. In this paper, we propose a novel curvature-based method to automatic Cobb angle measurement. The proposed Multi-task Vertebra Information Extraction network (namely MVIE-Net) is used to predict vertebra contour and keypoint confidence map simultaneously. And we pair the vertebral corner points based on the positional relationships contained in the vertebral contours and calculate the Cobb angle accordingly. The performance on the public AASCE Challenge dataset proves the efficiency of the proposed method. Experimental results on external datasets demonstrate the more generalizability of the proposed method.