Abstract:

In order to address the challenges caused by limited medical resources in large-scale Diabetic Retinopathy （DR） screening process， a dual-branch hybrid attention decision net （BiRAD-Net） for DR classification is proposed. The proposed network consists of feature extraction and classification two stages. In the feature extraction stage， a hybrid attention is introduced to suppress the noise， and feature grade decision network is designed to improve feature quality. In the feature classification stage， dual-branch classifiers and corresponding dual-branch loss are designed to alleviate the impact of insufficient data with five-category labels and enhance the classification accuracy. Furthermore， the transfer learning technology is applied in the training process， together with the above steps， to improve the accuracy of model and reduce the amount of training data. Experimental results on KAGGLE dataset indicate that BiRAD-Net shows excellent diagnostic ability for all the stages of DR and preforms better than other existing comparison methods.

Key words: computer application, deep learning, diabetic retinopathy（DR） classification, hybrid attention mechanism, feature grade decision net