基于改进U-Net模型和CBAM的脑肿瘤MRI图像分割  

MRI Image Segmentation of Brain Tumor Based on Improved U-Net Model and CBAM

作　　者：张晓倩 罗建 杨梅 金芊芊 朱熹 Zhang Xiaoqian;Luo Jian;Yang Mei;Jin Qianqian;Zhu Xi(School of Electronic Information Engineering,Xihua Normal University,Nanchong 637009)

机构地区：[1]西华师范大学电子信息工程学院,南充637009

出　　处：《现代计算机》2022年第16期60-66,共7页Modern Computer

基　　金：四川省教育厅重点项目(14ZA0123);西华师范大学英才科研基金项目(17YC157)。

摘　　要：由于脑肿瘤图像分割常出现缺少上下文信息和分割效率低的情况,本文采用对U-Net模型改进的方法来解决此问题。首先,在U-Net中引入改进的残差块,其作用是对网络进行深化,便于提取更多的特征信息;然后通过加入CBAM机制的方法,使得网络更加关注某些特征层和空间区域,抑制非病灶区域的特征,以提高肿瘤分割的准确性。本文利用MICCAI提供的公开磁共振成像MRI数据集对改进模型进行验证,并利用Dice系数和精准率对模型进行评估,在整体、核心和增强肿瘤区两者分别达到0.883、0.80、0.789和0.899、0.845、0.804。结果显示,修改后的模型对脑肿瘤图像的分割精度和准确率均有所提高。Due to the lack of context information and low segmentation efficiency in brain tumor image segmentation,this paper is based on a better U-Net model to solve this problem.Firstly,an improved residual block is introduced into U-Net to deepen the network and extract more feature information.Then,by adding CBAM mechanism,the network concentrates more on some feature layers and spatial regions,and suppresses the features of non-lesion areas to improve the accuracy of tumor segmentation.In this paper,the improved model was verified using the public magnetic resonance imaging MRI data set provided by MICCAI,and using the Dice coefficient and precision rate to evaluate the model.The overall,core and enhanced tumor areas reached 0.883,0.80,0.789 and 0.899,0.845,0.804,respectively.The results show that the segmentation accuracy and accuracy of the modified model for brain tumor images are improved.

关 键 词：深度学习 脑肿瘤分割 U-Net CBAM注意力机制 深度残差模块 

分 类 号：R739.41[医药卫生—肿瘤] TP391.41[医药卫生—临床医学] R445.2[自动化与计算机技术—计算机应用技术]