Automatic assessment of the Cobb angle can improve the efficiency of the clinical diagnosis of scoliosis. Existing automated scoliosis detection methods fail to deal with the small proportion of vertebrae in X-ray images, blurred vertebral boundaries, and local occlusion, thus limiting their clinical application. To overcome these problems, we propose the Dual Decoding Enhanced W-net (DDEW-net), a multi-task network with a dual decoder structure, which uses a global to local learning process. Specifically, the intermediate decoder performs the overall segmentation of the vertebrae in a deep supervision form, and its dense prediction result, in the form of a spatial attention map, enhances the information representation of the vertebral regions. The tail decoder, assisted by the intermediate decoder, performs local vertebral center point localization and vertebral inclination estimation. Extensive experiments on the public AASCE challenge dataset demonstrate that DDEW-net achieves better performance than the state-of-the-art methods.