MCMLNet: A Neural Network for Multiple Fundus Disease Classification Using Multi-Channel Mutual Learning Mechanism

With the aging of population and the change of people's lifestyle, fundus diseases have become the main irreversible cause of blindness. Computer-aided classification of multiple fundus diseases can improve the efficiency of diagnosis and prevent the further deterioration of fundus diseases. However, the classification approach faces great challenges due to the limited available data, complex types of fundus diseases, and subtle lesion features. To solve the above problems, we propose a neural network using multi-channel mutual learning mechanism. In this work, we propose Attention-Augmentation Block, which adjusts the learning focus of different channels by disease-grouping training on samples to make full use of the limited data. Furthermore, to enhance the feature representation capability of the model, the proposed method implements the mutual enhancement and fusion of multi-channel features through the mutual learning mechanism by designing the Multi-Channel Mutual Learning Unit. Experimental results show that our proposed method can achieve better classification performance of multiple fundus diseases than the state-of-the-art methods. The code of our paper is available in the attachment.