面向医学图像分割的双通道特征感知网络  

A dual-channel feature perception network for medical image segmentation

作　　者：武相虎 束鑫 范燕 黄树成 史金龙 WU Xiang-hu;SHU Xin;FAN Yan;HUANG Shu-cheng;SHI Jin-long(School of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212100,China;Development and Related Diseases of Women and Children Key Laboratory of Sichuan Province,Chengdu 610000,China)

机构地区：[1]江苏科技大学计算机学院,江苏镇江212100 [2]发育与妇儿疾病四川省重点实验室,成都610000

出　　处：《控制与决策》2025年第3期871-879,共9页Control and Decision

基　　金：江苏省研究生科研与实践创新计划项目(SJCX24_2541);发育与妇儿疾病四川省重点实验室开放课题项目(2023003);国家自然科学基金面上项目(62276118).

摘　　要：卷积神经网络(CNN)在医学图像分析领域得到了广泛应用,但是受其固定感受野的局限性,传统的CNN模型难以建立图像中的长距离依赖关系.Transformer通过自注意力机制能够建立图像全局视角下的信息依赖,拥有更强的序列建模能力.然而,Transformer难以捕获图像的局部细节特征.为了解决上述问题,提出一种基于CNN与Transformer的混合模型DC-TransNet,用于医学图像分割.首先,DC-TransNet采用双解码器结构建立图像局部和长距离依赖,以捕获局部和全局特征;然后,考虑到基于编码器-解码器结构的网络模型在不同深度提取到的特征图大小不一致,设计两种特征感知注意力机制CFP和SFP,以合理分配局部和全局特征的权重;最后,在多个医学数据集上进行实验.实验结果表明:DC-TransNet在2D医学图像单类别分割任务中取得了有竞争力的结果,mIoU和mDice等系数均得到了显著提升.Convolutional neural networks(CNNs)have been widely used in the field of medical image analysis.However,due to the limitation of its fixed receptive field,the traditional CNN model makes it difficult to establish long-distance dependencies in images.Transformer can establish the information dependence in the global perspective of the image through the self-attention mechanism and has stronger sequence modeling ability.However,Transformer makes it difficult to capture the local detailed features of images.To solve the above problems,a hybrid model DCTransNet based on the CNN and Transformer is proposed for medical image segmentation.DC-TransNet uses a dualdecoder structure to establish local and long-distance dependencies in the image and capture local and global features.Considering that the size of the feature maps extracted by the network model based on the encoder-decoder structure is inconsistent at different depths,we design two feature perception attention mechanisms,channel feature percep-tion attention(CFP)and spatial feature perception attention(SFP),to reasonably allocate the weight of local and global features.Experiments are conducted on multiple medical datasets and the results show that DC-TransNet is effective in 2D medical images.Competitive results are achieved in single-category segmentation tasks,and coefficients such as mIoU and mDice are significantly improved.

关 键 词：深度学习 卷积神经网络 TRANSFORMER 注意力机制 多尺度特征提取 医学图像分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]