基于多U-Net网络的脑胶质瘤分割算法研究  

Research on Glioma Segmentation Algorithm Based on Multi-U-Net Network

作　　者：刘宏 喻昕 蒋娟 伍胜 徐聪 乃科 张锦 LIU Hong;YU Xin;JIANG Juan;WU Sheng;XU Cong;NAI Ke;ZHANG Jin(College of Information Science and Engineering,Hunan Normal University,Changsha 410081,China;School of Computer and Communication Engineering,Changsha University of Science&Technology,Changsha 410114,China)

机构地区：[1]湖南师范大学信息科学与工程学院,湖南长沙410081 [2]长沙理工大学计算机与通信工程学院,湖南长沙410114

出　　处：《软件导刊》2024年第11期158-165,共8页Software Guide

基　　金：国家自然科学基金项目(61972055);湖南省自然科学基金项目(2021JJ30456,2021JJ30734);工业控制技术国家重点实验室开放课题(ICT2022B60);国防科技重点实验室基金项目(2021-KJWPDL-17)。

摘　　要：核磁共振影像分割对于脑肿瘤患者的治疗至关重要,但肿瘤形态多变、边界模糊等问题使得其边缘分割效果不佳。为解决以上问题,提出一种基于多U-Net网络的脑胶质瘤自动化分割算法MU-Net。首先,以U-Net为主干网络,在编码阶段设计残差空洞卷积模块作为短连接,增强编码特征长距离信息的连接,从而改善特征提取效果;其次,在网络跳跃连接处引入改进的高效通道注意力机制,同时使用平均池化和最大池化充分利用空间和通道信息,以提高分割准确度;最后,在经过改进高效通道注意力机制处理后的跳跃连接处设计多个不同深度的双输出U-Net作为编码与解码之间的纽带,以增强网络对不同尺度脑肿瘤的适应性。在BraTS2020数据集上进行大量实验,结果表明MU-Net算法对完整肿瘤、肿瘤核心和增强肿瘤的Dice系数分别为86.75%、77.76%和76.21%,与基准模型相比分别提升了2.6%、2.55%和2.41%,具有更好的分割效果。Magnetic resonance imaging segmentation is crucial for the treatment of brain tumor patients,but issues such as variable tumor morphology and blurred boundaries result in poor edge segmentation performance.To solve the above problems,a brain glioma automated seg⁃mentation algorithm MU Net based on multiple U-Net networks is proposed.Firstly,using U-Net as the backbone network,a residual dilated convolution module is designed as a short connection in the encoding stage to enhance the connection of long-distance information of encoded features,thereby improving the feature extraction effect;Secondly,an improved efficient channel attention mechanism is introduced at the network skip connections,while using average pooling and maximum pooling to fully utilize spatial and channel information to improve seg⁃mentation accuracy;Finally,multiple dual output U-Nets of different depths are designed at the skip connections processed by the improved efficient channel attention mechanism as links between encoding and decoding to enhance the network's adaptability to brain tumors of differ⁃ent scales.A large number of experiments were conducted on the BraTS2020 dataset,and the results showed that the MU Net algorithm had Dice coefficients of 86.75%,77.76%,and 76.21%for intact tumors,tumor cores,and enhanced tumors,respectively.Compared with the benchmark model,it improved by 2.6%,2.55%,and 2.41%,respectively,and had better segmentation performance.

关 键 词：脑肿瘤 核磁共振影像分割 多U-Net网络 残差模块 高效通道注意力机制 双输出U-Net 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]