基于注意力机制的脑肿瘤分割算法研究  

Research on segmentation algorithm of brain tumor based on attention mechanism

作　　者：戴昂 宋亚男[1] 徐荣华[1] 方俞泽 Dai Ang;Song Yanan;Xu Ronghua;Fang Yuze(School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学自动化学院,广东广州510006

出　　处：《无线互联科技》2023年第4期107-111,共5页Wireless Internet Technology

基　　金：广东工业大学高水平大学建设研究生教育创新计划项目,项目名称:实践创新能力培养模式研究,项目编号:2018J-GMS-09;广东省本科高校在线开放课程指导委员会研究课题,项目名称:多模态数据融合的学习分析与评价创新,项目编号:2022ZXKC143。

摘　　要：文章针对形状复杂、边界模糊的脑肿瘤难以实现精确分割的问题,提出一种基于卷积注意力机制和Transformer多头注意力机制的U型分割网络。文章首先设计了基于通道注意力和空间注意力的卷积模块,提高了模块对局部关键特征的提取能力;其次使用一种结构更精简的Transformer模块作为网络的瓶颈层,利用其多头注意力机制对全局特征进行充分感知;最后在BraTS 2021数据集上进行了实验。实验结果表明文章算法在增强肿瘤区域、肿瘤核心区域和整个肿瘤区域的Dice系数评分分别为87.51%,90.69%和93.47%,可以有效提高脑肿瘤分割精度。In this paper,a U-shaped segmentation network based on convolutional attention mechanism and Transformer multi-head attention mechanism is proposed to solve the problem that brain tumors with complex shapes and blurred boundaries are difficult to achieve accurate segmentation.Firstly,a convolutional module based on channel attention and spatial attention is designed to improve the module's ability to extract local key features.Secondly,a simpler Transformer module is used as the bottleneck layer of the network,and its multi-head attention mechanism is used to fully perceive the global features;Finally,an experiment is carried out on the BraTS 2021 dataset.The results show that the Dice coefficient scores of the algorithm in the enhancement tumor region,the tumor core region and the whole tumor region are 87.51%,90.69%and 93.47%respectively,which can effectively improve the accuracy of brain tumor segmentation.

关 键 词：脑肿瘤分割 混合网络 TRANSFORMER 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]