摘要
针对传统肝脏分割方法十分依赖医生的经验,并且分割过程耗时,易出错的现象,本文提出适用于临床情景中医学肝脏计算机断层扫描的分割方法。基于多尺度残差混合注意力U-Net将多尺度注意力机制模块引入U-Net网络。该模块可以抑制不相关的区域,从多个角度提取图像特征,并突出显示分割任务;在标准卷积层中添加残差结构可以有效地避免梯度爆炸并增加网络深度;使用混合空洞注意力常规层来替换“U”形网络的底部,以获得上下文信息,避免空间信息的丢失。试验结果表明:在LiTS17和SLiver07数据集上与其他方法相比,本文方法具有更好的性能和最高的分割精度。
In response to the challenges of traditional liver segmentation methods,which are highly dependent on doctors′experience and are time-consuming and prone to errors,a liver segmentation method suitable for clinical scenarios using medical liver CT images has been proposed.The multi-scale attention mechanism modules are incorporated into the U-Net network,specifically the multi-scale residual hybrid attention U-Net.These modules can suppress irrelevant regions,extract image features from multiple angles,and highlight the segmentation task.Incorporating a residual structure into the standard convolutional layer effectively avoids gradient explosion and increases network depth.Using a hybrid attenuation attention regular layer at the bottom of the“U”-shaped network helps obtain context information and prevents the loss of spatial information.Experimental results show that our method outperforms other methods on the LiTS17 and SLiver07 datasets and achieves the highest segmentation accuracy.
作者
金涛
王震
李昭蒂
JIN Tao;WANG Zhen;LI Zhaodi(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China;College of Computer and Control Engineering,Qiqihar University,Qiqihar 161006,China;School of Public Health,Qiqihar Medical College,Qiqihar 161006,China)
出处
《哈尔滨工程大学学报》
北大核心
2025年第3期529-539,共11页
Journal of Harbin Engineering University
基金
黑龙江省省属本科高校基本科研业务费项目(2024-KYYWF-0345).
关键词
神经网络
深度学习
语义分割
肝脏分割
医学图像
注意力机制
空洞卷积
neural network
deep learning
semantic segmentation
liver segmentation
medical images
attention mechanisms
atrous convolution
作者简介
金涛,男,副教授;通信作者:王震,男,助教,硕士,E-mail:wangzhen001513@qmu.edu.cn。
