Cobb Angle Rectification with Dual-Activated Linformer

As a current gold standard to measure spine deformity, Cobb angle estimation heavily relies on vertebrae localization. However, output angles may become extremely inaccurate when horizontal offset errors between predicted vertebrae and ground-truth ones appear large. Even though output angles are sometimes accurate, vertical offset errors remain huge. Recently, Transformers can capture complex relationships among tokens in many natural language processing applications. Therefore, to further reduce the errors from both keypoints and Cobb angles, we introduce a Cobb angle rectification mechanism based on Dual-Activated Linformer. And we integrate it into a one-stage keypoint estimator. Specifically, a Linformer fuses convolutional feature maps to capture concrete visual dependencies of vertebrae, and a Dual-Activated variant corrects keypoint parameters to learn abstract representations. Extensive experiments on anterior-posterior (AP) X-ray images from AASCE MICCAI 2019 Challenge show the effectiveness of our rectification mechanism.