LMD-UNet网络在脑肿瘤多模态MRI图像分割中的应用  

Application of LMD-UNET Network in Multi-Modal MRI Images Segmentation of Brain Tumors

作　　者：夏景明[1] 谈玲[2] 梁颖 Xia Jingming;Tan Ling;Liang Ying(School of Artificial Intelligence,Nanjing University of Information Science&Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics Ministry of Education,Nanjing University of Information Science&Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学人工智能学院,南京210044 [2]南京信息工程大学数字取证教育部工程研究中心,南京210044

出　　处：《中国生物医学工程学报》2023年第4期431-441,共11页Chinese Journal of Biomedical Engineering

基　　金：国家重点研发计划科技创新2030—“新一代人工智能”重大项目(2021ZD0112200);江苏省产学研基金(BY2022459)。

摘　　要：UNet网络中编解码器对应的特征图之间存在语义鸿沟,其双卷积层无法学习多尺度信息,造成部分特征信息丢失,影响MRI图像分割效果。针对这一缺陷,本研究提出一种新的图像分割网络局部残差融合多尺度双分支网络LMD-UNet。在编码流程,网络采用局部特征残差融合密集块和多尺度卷积模块,扩大影像感受野并优化底层视觉特征的传播;在解码流程,网络采用双分支卷积的方式生成新的高级语义特征,以此来重建编码路径中损失的信息。利用公开脑肿瘤数据集BraTs的335例病例做分割实验,并将分割结果与现阶段主流分割网络UNet进行对比。结果显示,LMD-UNet模型的Precision、Dice、95%HD、Recall等4项客观评价指标分别达到0.933、0.921、0.702和0.966,相较于UNet,对应指标分别提升了6.3%、5.7%、1.8%和6.1%。研究表明,LMD-UNet能够实现更精细的脑肿瘤图像分割。此外,所提出的方法对于细节部分边缘轮廓的分割也有较好的效果,能够为脑肿瘤诊断和手术提供保障。There is a semantic gap among the feature maps corresponding to the codec in the UNet network,and its dual roll integration layer cannot learn multi-scale information,resulting in the loss of some feature information,which affects the MRI image segmentation effect.To solve this problem,this paper proposed a new image segmentation network local residual fusion multi-scale dual branch network LMD-UNet.In the coding process,the network used local feature residuals to fuse dense blocks and multi-scale convolution modules to expand the receptive field of images and optimize the propagation of underlying visual features;and in the decoding process,the network used double branch convolution to generate new high-level semantic features to reconstruct the information lost in the coding path.For segmentation experiments,335 cases of the public brain tumor dataset BraTs were used,and the segmentation results were compared with U-Net that is currently a mainstream segmentation network.Experimental results showed that the four objective evaluation indexes of LMD-UNet model,precision,dice,95%HD and recall reached 0.933,0.921,0.702 and 0.966 respectively.Compared to U-Net,the corresponding indicators increased by 6.3%,5.7%,1.8%,and 6.1%,respectively,which indicated that LMD-UNet achieved more precise segmentation of brain tumor images.Meanwhile,the proposed method also showed a good performance in the edge contour segmentation for the detail part,which prospectively provided guarantee for the diagnosis of brain tumor and the surgery.

关 键 词：肿瘤图像分割 多尺度卷积 密集块 双分支卷积 

分 类 号：R318[医药卫生—生物医学工程]