Automatic Segmentation of Lumbar Vertebra Anatomical Region Based on Hybrid Swin-Transformer Network

Lumbar vertebra anatomical region segmentation is crucial in an automated spine processing pipeline. However, the variations in anatomy and severe lack of publicly available data hinder designing automated algorithms for lumbar segmentation. Convolutional neural networks (CNN) have been recently the de facto standard in medical image segmentation. Although capable of capturing global feature, the inductive bias of locality and weight sharing of CNN limit its ability of modeling long-distance dependency. In this paper, we propose a novel hybrid architecture for lumbar vertebrae anatomical region segmentation, which hybridizes Swin-Transformer and multi-scale atrous convolution as encoder to produce hierarchical representations from global and local features. Besides, a feature coupling module is introduced to incorporate features from transformer block to convolution block in the channel and spatial dimensions. To address the problem of lack of data, we raise a new dataset called LumASe containing a total of 663 lumbar vertebrae annotated at voxel level by 3 physicians. Extensive experimental results on LumASe demonstrate the proposed algorithm achieves better results compared to the state-of-the-art methods for the medical image segmentation. The data is publicly available at: https://zenodo.org/record/7181338 - .Y1S1Rey-sUE