基于注意力机制与特征融合的耳诊图像五脏反射区分割  

Mapping Zone Segmentation of Five Organs in Auricular Images Based on the Attention Mechanism and Feature Fusion

作　　者：冯跃 梁惠珠 徐红 林卓胜 张双胜[2] 胡敏儿 FENG Yue;LIANG Hui-zhu;XU Hong;LIN Zhuo-sheng;ZHANG Shuang-sheng;HU Min-er(Faculty of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,China;Jiangmen Central Hospital,Jiangmen 529030,China)

机构地区：[1]五邑大学智能制造学部,广东江门529020 [2]江门市中心医院,广东江门529030

出　　处：《五邑大学学报（自然科学版）》2022年第3期35-43,共9页Journal of Wuyi University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61901304);广东省普通高校重点领域专项项目(2021ZDZX1032);广东省国际及港澳台高端人才交流专项(2020A1313030021);五邑大学科研项目(2018TP023)。

摘　　要：计算机技术辅助分析中医耳诊图像时,五脏反射区在耳部图像中为小目标区域且边缘模糊,存在难以分割等问题.为此,本文提出一种多视图空间注意力与特征融合分割模型.在U-Net编码端设计自适应平均池化分支,即将输入图像的底层特征与不同编码层的高层特征融合,补偿U-Net模型信息损失;编码端和解码端设计多视图空间注意力模块进行连接,模块通过对特征进行不同尺寸划分,实现目标区域有效信息提取,兼顾全局与局部位置关系;为了捕获更多目标区域细节信息,解码端最后输出层设计多尺度特征融合模块,融合不同感受野下的输出特征.实验结果表明,所提模型在两个耳诊图像数据集上的平均Jaccard系数、Dice系数、ASSD分别为48.86%、65.40%、1.57像素和47.31%、63.72%、2.30像素,比U-Net提高了8.91%、8.56%、1.13像素和9.39%、9.44%、1.83像素,模型在两个数据集上表现出较好的分割性能.It is difficult for five-organ mapping zones to be segmented because they are small target regions with blurred edges in auricular images.A multi-view spatial attention and feature fusion semantic segmentation model is presented for solving this problem.Based on U-Net,an adaptive average pooling branch is designed on the encoder to fuse the low-level features of the input image with the higher-level features of different encoder layers to compensate for information loss.The encoding end and decoding end are connected by a multi-view spatial attention module.The module can extract effective information of the target region by dividing features into different sizes,and give consideration to global and local location relations.In order to capture more details of target regions,a multi-scale feature fusion module is designed at the last output layer of the decoding end to fuse the output features in different receptive fields.The experimental results show that the average Jaccard coefficient,Dice coefficient and ASSD of the proposed model are 48.86%,65.40%,1.57 pixels and 47.31%,63.72%,2.30 pixels,respectively,on the two ear diagnosis image data sets.Compared with U-Net,the segmentation performance is improved by 8.91%,8.56%,for 1.13 pixels and 9.39%,9.44%,for 1.83 pixels,respectively.The model shows good segmentation performance on the two datasets.

关 键 词：耳诊 图像分割 深度学习 U-Net 注意力机制 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]