Spine X-ray Image Segmentation Based on Transformer and Adaptive Optimized Postprocessing

Accurate segmentation of vertebral blocks in X-ray images of the whole spine is necessary for intelligent diagnosis of spinal diseases. However, the vertebral block in X-ray spine image has similar appearance with background, which cause a huge challenge to segmentation performance. Even though the classical Unet network and existing popular methods have been proved as effective approaches in medical image segmentation, they still create some conglutination between vertebral blocks in spine X-ray images. To overcome this problem, we propose a novel spine X-ray image segmentation method based on transformer and adaptive optimized postprocessing to improve segmentation performance. Firstly, we adopt the Segmenter network based on Transformer framework to obtain initial result. Then, an adaptive post-processing optimization strategy is proposed to optimize segmentation results. Finally, experiments on AASCE 2019 challenge dataset show that our model can gain best performance compared with existing methods.