RAU-Net: U-Net network based on residual multi-scale fusion and attention skip layer for overall spine segmentation

Spine segmentation is necessary for the clinical quantitative analysis of computed tomography (CT) images and plays an important role in the early diagnosis and treatment of spine diseases. However, because of the different fields of view of sagittal CT, these images show different shapes and sizes of vertebrae, unclear vertebral boundaries, and different image scales which greatly complicate the segmentation of the spine. To solve this problem, we propose a new deep learning method for segmenting the spine. For this algorithm, we first proposed a residual feature pyramids block for capturing and fusing multi-scale information. For fusing shallow and deep features, we then propose an attention skip layer structure for suppressing the reuse of redundant information. Finally, we use a joint loss function to optimize the segmentation results and achieve the effect of clear segmentation edges. Through the combination of these three techniques, our method achieves efficient and accurate spinal segmentation. The experimental results show that our model has a good performance in spine segmentation. In particular, our achieve the Dice of 0.8973, the Hausdorff distance of 77.6277 on the VerSe 2019 dataset and the Dice of 0.8626, the Hausdorff distance of 82.6170 on the VerSe 2020 dataset.