基于加性注意力的前列腺MR图像分割算法  

Prostate MR image segmentation algorithm based on additive attention

作　　者：章子烽 薛仁政[1,2] 叶宝丰 吴乾龙 徐海强 ZHANG Zi-feng;XUE Ren-zheng;YE Bao-feng;WU Qian-long;XU Hai-qiang(College of Computer and Control Engineering,Qiqihar University,Heilongjiang Qiqihar 161006,China;Key Laboratory of Big Data Network Security Detection and Analysis,Heilongjiang Qiqihar 161006,China)

机构地区：[1]齐齐哈尔大学计算机与控制工程学院,黑龙江齐齐哈尔161006 [2]黑龙江省大数据网络安全检测分析重点实验室,黑龙江齐齐哈尔161006

出　　处：《齐齐哈尔大学学报(自然科学版)》2025年第3期31-36,共6页Journal of Qiqihar University(Natural Science Edition)

基　　金：黑龙江省教育厅基本科研业务专项(145209127)。

摘　　要：前列腺核磁共振图像组织边缘模糊,难以准确分割。为此提出一种基于编解码器结构的前列腺图像分割模型(EAAU-Net)。基于细节增强卷积构建DE-ResNet50作为特征提取器,在编码器末端引入加性注意力模块建模全局上下文。跨尺度边缘增强模块从浅层特征中提取跨尺度边缘信息进行对齐,增强细节。解码器级联多个像素注意力特征融合模块,有选择地学习高层语义信息和细节信息,降低不同层级语义差异。Prostate magnetic resonance(MR)images have blurred tissue edges,making accurate segmentation challenging.Therefore,this study proposes a prostate image segmentation model based on an encoder-decoder structure(EAAU-Net).DE-ResNet50 is constructed as a feature extractor based on detail-enhanced convolution,and an additive attention module is introduced at the end of the encoder to model the global context.The cross-scale edge enhancement module extracts cross-scale edge information from shallow features for alignment,enhancing details.The decoder cascades multiple pixel attention feature fusion modules to selectively learn high-level semantic information and detail information,reducing semantic differences at different levels.

关 键 词：前列腺图像 深度学习 图像分割 注意力机制 跨尺度边缘增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]