多分辨率融合输入的U型视网膜血管分割算法

Multi-resolution Fusion Input U-shaped Retinal Vessel Segmentation Algorithm

针对视网膜血管拓扑结构不规则、形态复杂和尺度变化多样的特点,该文提出一种多分辨率融合输入的U型网络(MFIU-Net),旨在实现视网膜血管精准分割。设计以多分辨率融合输入为主干的粗略分割网络,生成高分辨率特征。采用改进的ResNeSt代替传统卷积,优化血管分割边界特征;将并行空间激活模块嵌入其中,捕获更多的语义和空间信息。构架另一U型精细分割网络,提高模型的微观表示和识别能力。一是底层采用多尺度密集特征金字塔模块提取血管的多尺度特征信息。二是利用特征自适应模块增强粗、细网络之间的特征融合,抑制不相关的背景噪声。三是设计面向细节的双重损失函数融合,以引导网络专注于学习特征。在眼底数据用于血管提取的数字视网膜图像(DRIVE)、视网膜结构分析(STARE)和儿童心脏与健康研究(CHASE_DB1)上进行实验,其准确率分别为97.00%,97.47%和97.48%,灵敏度分别为82.73%,82.86%和83.24%,曲线下的面积(AUC)值分别为98.74%,98.90%和98.93%。其模型整体性能优于现有算法。

Considering the characteristics of irregular retinal blood vessel topology,complex morphology and diverse scale changes,a Multi-resolution Fusion Input U-Netword(MFIU-Net)is proposed to achieve accurate segmentation of retinal blood vessels.A rough segmentation network based on multi-resolution fusion input is designed to generate high-resolution features..The improved ResNeSt is used to replace the traditional convolution to optimize the boundary features of blood vessel segmentation,and the parallel spatial activation module is embedded to capture more semantic and spatial information.Another U-shaped fine segmentation network is constructed to improve the microscopic representation and recognition ability of the model.Firstly,the multi-scale dense feature pyramid module to extract the multi-scale feature information of blood vessels is adopted at the bottom layer.Secondly,the feature adaptive module is used to enhance the feature fusion between coarse and fine networks to suppress irrelevant background noise.Thirdly,a detail-oriented double loss function fusion is designed to guide the network to focus on learning features.Experiments are carried out on the fundus data Digital Retinal Images for Vessel Extraction(DRIVE),STructured Analysis of the REtinal(STARE)and Child Heart and Health Study(CHASE_DB1),the accuracy rates are 97.00%,97.47%and 97.48%,the sensitivity is 82.73%,82.86%and 83.24%,and the Area Under Cure(AUC)values are 98.74%,98.90%and 98.93%,respectively.The overall performance of its model is better than that of existing algorithms.