关键词: 脑肿瘤图像分割, UNet++, MCAM, CA注意力机制, Mish激活函数, SME, 焦点Dice损失函数

Abstract: Aiming at the problems of blurred edge segmentation and low segmentation accuracy of computer-aided brain tumor images, an improved nested UNet++ brain tumor image segmentation algorithm was proposed. Firstly, the MCAM module is designed to replace the feature extraction part of the original UNet++. The CA attention mechanism is used to focus on position information in different directions to improve feature extraction capabilities. The Mish activation function is used to replace the ReLU activation function to prevent gradient disappearance and improve segmentation. accuracy and generalization ability; then add the SME module after feature extraction for squeezing and excitation to enhance the receptive field of the feature map to enhance the learning ability of tumor features; finally, use the focus Dice loss function to focus on fuzzy samples segmentation, thereby improving the problem of blurred- DOI:10.3969/j.issn.1000-1565.2025.04.008改进UNet++模型的脑肿瘤图像分割算法付豪,张振利,陈源(江西理工大学 电气工程与自动化学院,江西 赣州 341000)摘 要:针对计算机辅助脑肿瘤图像边缘分割模糊、分割精度不高的问题,提出了一种改进的嵌套UNet++脑肿瘤图像分割算法.首先,设计MCAM(Mish coordinate attention module)模块代替原UNet++的特征提取部分,嵌入坐标注意力机制(coordinate attention,CA)关注不同方向上的位置信息以增强特征提取能力,使用Mish激活函数替换ReLU激活函数防止出现梯度消失,提高脑肿瘤图像分割精度和泛化能力;其次,在特征提取后加入SME(squeeze Mish excitation)模块进行挤压和激励,扩大特征图的感受野以增强对肿瘤特征的学习能力;最后,利用焦点Dice损失函数关注模糊样本的分割,从而改善脑肿瘤图像边缘分割模糊的问题.提出的算法在Figshare数据集上进行仿真实验,实验结果表明,在均值交并比(MIoU)、类别平均像素准确率(MPA)、骰子系数(Dice)和豪斯多夫距离(Hausdorff distance,HD)评估指标上分别达到83.26%、81.91%、86.45%和18.57 mm.与3DUNet、Swin-UNet、DD-UNet、LRAE-UNet和AI-UNet等算法进行对比,证明提出的算法分割效果更优.关键词:脑肿瘤图像分割;UNet++;MCAM;CA注意力机制;Mish激活函数;SME;焦点Dice损失函数中图分类号:TP391.41 文献标志码:A 文章编号:1000-1565(2025)04-0398-10Brain tumor image segmentation algorithm based on an improved UNet ++ modelFU Hao, ZHANG Zhenli, CHEN Yuan(School of Electrical Engineering and Automation, Jiangxi University of Science and Technology, Ganzhou 341000, China)Abstract: Aiming at the problems of blurred edge segmentation and low segmentation accuracy of computer-aided brain tumor images, an improved nested UNet++ brain tumor image segmentation algorithm was proposed. Firstly, the MCAM module is designed to replace the feature extraction part of the original UNet++. The CA attention mechanism is used to focus on position information in different directions to improve feature extraction capabilities. The Mish activation function is used to replace the ReLU activation function to prevent gradient disappearance and improve segmentation. accuracy and generalization ability; then add the SME module after feature extraction for squeezing and excitation to enhance the receptive field of the feature map to enhance the learning ability of tumor features; finally, use the focus Dice loss function to focus on fuzzy samples segmentation, thereby improving the problem of blurred- 收稿日期:2024-10-17;修回日期:2025-04-27 基金项目:国家自然科学基金项目(62363013) 第一作者:付豪(2000—),男,江西理工大学在读硕士研究生,主要从事计算机医学图像处理方向研究.E-mail:1324940506@qq.com 通信作者:张振利(1976—),男,江西理工大学副教授,主要从事人工智能、检测技术与控制方向研究.E-mail:47717770@qq.com第4期付豪等:改进UNet++模型的脑肿瘤图像分割算法河北大学学报(自然科学版) 第45卷edge segmentation of brain tumor images. The proposed algorithm was simulated on the Figshare dataset. The experimental results showed that it achieved 83.26%, 81.91%, 86.45% and 18.57mm in MIoU, MPA, Dice and Hausdorff evaluation indicators respectively. Comparison with algorithms such as 3DUnet, Swin-UNet, DD-UNet, LRAE-UNet and AI-UNet proves that the proposed algorithm has better segmentation effect.

Key words: brain tumor image segmentation, UNet++, MCAM, coordinate attention mechanism, Mish activation function, SME, focal Dice loss function