An improved U-Net network for medical image segmentation

Abstract

In many computer-aided spinal imaging and disease diagnosis, automating the segmentation of the spine and cones from CT images is a challenging problem. Therefore, in this paper, we propose a triple channel expansion attention segmentation network based on U-Net for spinal CT images. We design a triple channel expansion attention to solve the problem of low accuracy caused by the loss of important feature information in the downsampling process of ordinary convolution, which uses different sizes of convolution set kernels to extract different features. Then through this attention, we output a feature image for each layer of the down-sampling, and finally skip connection with it during the up-sampling. Finally, many experimental results on VerSe 2019 and VerSe 2020 datasets show that our proposed network is superior to other prior art segmentation networks.