SPA-RESUNET: Strip Pooling Attention Resunet for Multi-Class Segmentation of Vertebrae and Intervertebral Discs

Automated multi-class segmentation of vertebrae and intervertebral discs for volumetric magnetic resonance image(MRI) can help the diagnosis and the treatment of many spinal diseases. However, existing methods based on fully convolutional network mostly stacked local convolution and pooling operations, and thus failed to capture the long-range dependencies for spine segmentation. In this paper, we propose a novel and computation-efficient residual U-Net with strip-pooling attention mechanism (SPA-ResUNet) for the multi-class segmentation of vertebrae and intervertebral discs. Due to the effectiveness of skip connections in U-Net, the strip pooling attention module is adopted in the skip connections to learn the long-range dependencies for improving the segmentation performance of the top structures and sacral crest of spine. Our SPA-ResUNet was implemented on the MRSpineSeg Challenge dataset. Experiments show that our SPA-ResUNet achieves impressive performance with mean Dice similarity coefficients of 88.14% and 87.88% in all 19 spinal structures on two test datasets, respectively. The visual segmentation results also show that our method achieves superior segmentation in top structures and sacral crest of spine. The proposed SPA-ResUNet has high potential in aiding the diagnosis and the treatment of spinal diseases.