摘要
糖尿病性视网膜病变(diabetic retinopathy,DR)是糖尿病在发病过程中影响视网膜的症状。针对模型下采样过程中特征提取DR图像微动脉瘤等病灶区域信息丢失问题,提出了一种DenseNet融合残差结构的模块。该模块首先连接两个连续的dense block,然后利用残差结构对特征信息求和,并行融合处理特征图像信息,以防止有效特征信息的丢失,最后残差连接两个含有dropout的卷积块,抑制过拟合现象。针对以往卷积操作中未对病变区域的特征图通道加权的问题,提出了一种SeNet融合残差结构的模块。该模块首先连接SeNet,把全局平均池化和全局最大池化的特征信息相加,以提高有效通道信息的利用率,然后通过Conv1×1的残差方式来保证特征图信息的完整性。基于以上两个模块的设计,提出了一种DenseNet和SeNet融合残差结构的DR分类方法。该模型在APTOS2019数据集上的精确度达到89.8%,特异性达到97.0%,在Messidor-2数据集上的精确度达到78.8%,特异性达到91.9%,能够有效地提高视网膜图像病变程度的分类能力。
Diabetic retinopathy is the symptom of diabetes affecting the retina during the onset of diabetes.Aiming at the problem of information loss of lesion areas such as microimages during model downsampling,this paper proposed a module of DenseNet fusion residual structure.This module firstly connected two consecutive dense blocks,then sumed the feature information using the residual structure,and processed the feature image information in parallel to prevent the loss of effective feature information.Finally,the residual connected the two convolution blocks containing drop out to suppress the overfitting phenomenon.To solve the problem of the channel weighting of the feature graphs of lesion areas in previous convolution operations,this paper proposed a module of SeNet fusion residue structure.This module firstly connected SeNet,added the feature information of global average pooling and global maximum pooling to improve the utilization of effective channel information,and then ensured the integrity of feature graph information through the residual mode of conv1×1.Based on the design of the above two modules,this paper proposed a DR classification method of DenseNet and SeNet fusion residue structures.The model achieves 89.8%precision and 97.0%specificity on the APTOS2019 dataset,78.8%accuracy and 91.9%specificity on the Messidor-2 dataset,which can effectively improve the classification ability of the degree of retinal lesions.
作者
宋鹏飞
吴云
Song Pengfei;Wu Yun(State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;College of Computer Science&Technology,Guizhou University,Guiyang 550025,China)
出处
《计算机应用研究》
CSCD
北大核心
2024年第3期928-932,950,共6页
Application Research of Computers
基金
贵州省科技计划资助项目(黔科合基础-ZK[2022]一般119)
贵州大学研究生创新人才计划项目
糖尿病视网膜图像分割分级模型研究的创新型人才培养实践。
