基于瓶颈残差注意力机制U-net的肝脏肿瘤分割

医学CT图像中含有的大量噪声以及肝脏肿瘤大小不均一、位置因人而异且与相邻器官较相似的组织密度等都造成肝脏肿瘤分割困难。现有传统全卷积神经网络（FCN）方法，通过为输入CT图像中每个像素分配类别标签来实现肝脏肿瘤分割，但在分割精度上仍会出现小目标漏检或目标边界分割模糊的问题。针对这些问题，本文提出一种瓶颈残差注意力机制U-net（BRA U-net）的肝脏肿瘤分割方法，通过引入瓶颈残差模块可在非常深的网络中大幅减少计算量的同时解决梯度消失问题，此外堆叠的注意力模块可以增大有效特征的比重。本文在公共MICCAI2017肝肿瘤分割数据集上测试了该框架，戴斯相似性系数值达到0.788，高于其他对比分割网络，并利用3D-IRCADb数据集来验证该方法的一般性，结果表明本文方法分割效果良好，能够为临床诊断提供可靠依据。 

Segmentation of Liver Tumors Based on Bottleneck Residual Attention Mechanism U-net

The segmentation of liver tumors is difficult due to the large amount of noise contained in medical CT images and the uneven size, location and tissue density of liver tumors similar to adjacent organs. The existing traditional full-convolutional neural network (FCN) method can achieve liver tumor segmentation by assigning category labels to each pixel in the input CT image, but the problems of small target missing or fuzzy target boundary segmentation still occur in the segmentation accuracy. In order to solve these problems, we propose a bottleneck residual attention mechanism U-net (BRA U-net) segmentation method for liver tumors. By introducing bottleneck residual modules, we can greatly reduce the computational load in very deep networks while solving the gradient disappearance problem. In addition, stacked attention modules can increase the proportion of effective features. In this paper, the proposed framework was tested on the public MICCAI2017 liver tumor segmentation dataset, and the dice similarity coefficient value reached 0.788, higher than other comparative segmentation networks. The generality of the proposed method was verified by using the 3D-IRCADb dataset. The results show that the proposed method has good segmentation effect and can provide a reliable basis for clinical diagnosis.