基于生成对抗网络与双注意力的糖网分类方法

Classification of Diabetic Retinopathy Using Generative Adversarial Networks and Dual Attention Network

针对在糖尿病视网膜病变分类过程中,因为数据集不均衡、类间特征相似、类内又存有差异,从而导致最终分类准确率不高的问题,提出了一种结合生成对抗网络与双注意力的分类方法AIDnet。首先,在ACGAN网络后加入转置卷积进行改进,生成轻度NPDR、重度NPDR、 PDR的图像平衡数据集;其次,在InceptionV3网络的基础上加入双注意力机制(DAM),在减少计算开销的同时提升性能;最后,利用焦点损失函数增加难以识别病变的权重,减少易识别病变的权重,高效提取DR图像的细节特征。实验结果表明,AIDnet网络在Kaggle数据集上的自动分类准确率为89.53%,敏感度为82.45%,特异性为93.26%;在Messidor2上的准确率达到90.31%,敏感度达到89.28%,特异性达到93.31%。较其他分类方法而言,AIDnet分类效果良好,有助于提高糖尿病视网膜病变的分类准确率。

In the classification process of diabetes retinopathy, the classification accuracy is not high due to uneven data sets, similar characteristics between classes, and differences within classes. In this paper, AIDnet, a method combining generative adversarial network and dual attention is proposed. Firstly, transposed convolution was added to the original ACGAN network to improve and generate mild NPDR, severe NPDR and PDR images, and the data set was balanced. Secondly, the double attention mechanism(DAM) was added on the basis of InceptionV3 structure, which reduced the computational cost and improves performance. Finally, the focal loss function was used to increase the weight of difficult-to-identify diseases and reduce the weight of easy-to-identify diseases, which efficiently extracted the detailed features of DR images. The experimental results on the Kaggle dataset show that the automatic classification accuracy of DR based on AIDnet network is 89.53%, the sensitivity is 82.45%, and the specificity is 93.26%, accuracy on the Messidor2 dataset is 90.31%, the sensitivity is 89.28%, and the specificity is 93.31%. Compared with other methods, the classification effect is better, which helps to improve the classification accuracy of diabetic retinopathy.