SATR: A Structure-Affinity Attention-Based Transformer Encoder for Spine Segmentation

In digital histopathology, spine segmentation on ultrasound images plays a vital role, especially as a pre-processing filter to measure spine deformity and diagnose scoliosis automatically. This segmentation task remains challenging owing to the lack of consideration of high spatial correlation for different bone features. In this paper, in order to encode the rich prior knowledge regarding their structural attributes and spatial relationships, we propose a novel structure-affinity attention-based transformer encoder (SATR) to segment spine. It exploits the hierarchical architecture to output multi-scale feature representations. Meanwhile, the constraint on spine structural information enhances the feature usability of the network and consequently improves the segmentation accuracy. The comparative experiments verify that SATR achieves promising performance on spine segmentation as compared with other state-of-the-art candidates, which makes it conveniently replace the backbone networks for intelligent scoliosis assessment.