基于改进U-net的医学图像分割模型  

Medical Image Segmentation Model Based on Improved U-net

作　　者：杨永谱 甘海涛 夏薇 杨智 叶志伟[1] YANG Yongpu;GAN Haitao;XIA Wei;YANG Zhi;YE Zhiwei(School of Computer Science,Hubei Univ.of Tech.,Wuhan 430068,China;Department Wuhan Children’s Hospital(Wuhan Maternal and Child Healthcare Hospital),Tongji Medical College,Huazhong Univ.of Sci.&Tech.,Wuhan 430070,China)

机构地区：[1]湖北工业大学计算机学院,湖北武汉430068 [2]华中科技大学同济医学院附属武汉儿童医院(武汉市妇幼保健院),湖北武汉430070

出　　处：《湖北工业大学学报》2025年第1期49-54,共6页Journal of Hubei University of Technology

基　　金：湖北工业大学高层次人才基金(GCRC2020016);湖北省自然科学基金(2021CFB282);省部共建生物催化与酶工程国家重点实验室开放基金项目(SKLBEE2021020和SKLBEE2020020)。

摘　　要：U-net模型在医学图像分割中取得了很多的发展,但是其忽略了低层视觉特征和高层语义特征之间的差距,限制了网络的特征提取能力。针对这个问题,首先设计了一个特征增强模块,以更大的感受野来增强网络的特征提取能力,并使用跳跃连接将网络中低级特征与高级特征融合。同时引入三重注意力机制来融合空间维度和通道维度的特征,从而达到增强有用信息的目的。最后,设计了一种新的损失函数,引入二元交叉熵和Dice损失函数,来抑制数据集中类别不平衡的问题。该模型在BraTs数据集上进行了实验,取得了0.8592的平均Dice。实验结果表明,该模型性能优于其他常用的分割网络。U-net model has made a lot of progress in medical image segmentation,but it ignores the gap between low-level visual features and high-level semantic features,which limits the feature extraction ability of the network.To solve this problem,a feature enhancement module is designed to enhance the feature extraction capability of the network with a larger receptive field,and a jump connection is used to fuse low-level features with high-level features in the network.At the same time,the triple attention mechanism is introduced to integrate the features of spatial dimension and channel dimension so as to enhance the useful information.Finally,a new loss function was designed,which introduced binary cross entropy and Dice loss function to suppress the category imbalance in the data set.The model was tested on the BraTs data set and the average Dice of 0.8592 was achieved.Experimental results show that the performance of this model is better than other common segmentation networks.

关 键 词：医学图像分割 特征增强 注意力机制 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]