MTRA-CNN: A Multi-Scale Transfer Learning Framework for Glaucoma Classification in Retinal Fundus Images

Abstract

In the diagnosis of glaucoma based on deep learning, it is more meaningful for ophthalmologist to make a graded diagnosis using fundus images to reflect the degree of disease. Such multi-classification diagnosis tasks require a larger amount of data to enable the neural network to extract more feature information, while it is very difficult in medical field. To address these issues, we propose a novel Multi-Scale Transfer Learning framework (MTRA-CNN) which can make a graded diagnosis effectively based on graded glaucoma dataset with small volume: (1) Aiming at the problem of small data volume, we innovatively apply multi-stage transfer learning technique in glaucoma, in which the advantage of different types of datasets such as ImageNet and Ocular disease intelligent recognition (ODIR) are taken. (2) We innovatively combine the Residual attention(RA) block, a functional module designed specifically for fundus images, with transfer learning techniques and pre-trained it on the ODIR dataset to further improve the network’s upper-level feature extraction ability of glaucoma fundus images. (3) To achieve better diagnostic performance, we propose a novel multi-scale transfer learning method by combining the two-stage transfer learning with one-stage transfer learning. To validate our framework, we conducted experiments on the private glaucoma 4 classification dataset, and the results show that the accuracy of our method is 86.8%, compared to the method without any transfer learning, the accuracy of our experiments improved by 19.79%.

Keywords

• Multi-scale transfer learning
• graded glaucoma diagnosis
• deep CNNs
• upper-level feature
• bottom-level feature