An effective U-Net and BiSeNet complementary network for spine segmentation

Abstract

Spine segmentation with MRI plays an important role in the diagnosis and treatment of various spinal diseases, especially for multi-class segmentation of different vertebrae and intervertebral discs. However, due to the similarity between classes and differences within classes of spinal images, it is still a challenge to realize accurate multi-class fusion segmentation of different vertebrae and intervertebral discs. In this paper, we propose an effective U-Net and BiSeNet complementary network for spine segmentation to fully exploit their advantages in feature extraction. The network consists of a residual U-Net, a spatial feature extractor, and a feature fusion extractor. The residual U-Net selects the U-Net framework based on ResNet34, aiming at fully extracting the features of vertebrae and intervertebral discs in MRI. The spatial feature extractor takes the strip pooling (SP) blocks to replace the spatial extraction path in BiSeNet network, aiming at capturing long-range relations among the vertebrae and intervertebral discs and the interconnections of same class. The feature fusion extractor adopts the attention refinement modules (ARM) to aggregate the features from the residual U-Net and the spatial feature extractor, aiming at guiding the network to learn the distinguishing features between vertebrae and intervertebral discs by attention mechanism and making full use of the complementary features of the two paths. Experiments on MRI of 172 subjects show that our proposed network achieves impressive performance with mean Dice similarity coefficients of 82.802 % in all spinal structures. The proposed method has great value in the diagnosis and treatment of spinal diseases.