基于融合中间特征网络的视盘和视杯联合分割  

Joint Segmentation of Optic Disc and Cup Based on Middle Feature Fusion Network

作　　者：刘哲夏 李峰[1] 江旻珊[1] LIU Zhexia;LI Feng;JIANG Minshan(School of Optical-electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《控制工程》2024年第7期1272-1279,共8页Control Engineering of China

基　　金：国家自然科学基金资助项目(61905144)。

摘　　要：针对视盘和视杯联合分割中视杯分割精度较差的问题,提出了一种融合编码与解码中间特征的U型网络(encode-decode middle feature fusion U-Net,EMFF-Net)。EMFF-Net使用预训练的ResNet34作为编码结构,在编码结构后加入密集空洞卷积和金字塔池化模块以产生复合感受域的特征,并使用交叉注意力连接替换U型网络结构中的跳跃连接。交叉注意力连接融合了编码特征与解码特征,通过通道注意力模块和空间注意力模块提取融合特征的信息用于强化解码特征,减小了解码特征与编码特征的语义沟壑。强化后的解码特征与编码特征再次融合后,通过解码结构输出视盘和视杯的联合分割结果。实验结果表明,与其他常用的分割方法相比,EMFF-Net的视盘和视杯联合分割效果较好,视杯分割性能有明显提升。To solve the problem of poor accuracy of optic cup segmentation in the joint segmentation of optic disc and optic cup,an encode-decode middle feature fusion U-Net(EMFF-Net)is proposed.EMFF-Net uses pre-trained ResNet34 as the coding structure.The dense atrous convolution and pyramid pooling modules are added to the encoded structure to generate the features of the complex receptive field.The cross attention link is used to replace the skip link in the U-Net.The cross attention link combine encoding features and decoding features,and extracts the information of fusion features through channel attention module and spatial attention module to strengthen decoding features and reduce the semantic gap between decoding features and encoding features.After the enhanced decoding features and encoding features are fused again,the joint segmentation results of optic disc and optic cup are output through the decoding structure.The experimental results show that,compared with other common segmentation methods,EMFF-Net has better joint segmentation effect of optic disc and optic cup,and significantly improves the performance of optic cup segmentation.

关 键 词：视盘视杯分割 特征融合 EMFF-Net 深度学习 交叉注意力连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]