SAE-wAMD: A Self-Attention Enhanced convolution neural network for fine-grained classification of wet Age-Related Macular Degeneration using OCT Images

Abstract

Wet Age-related Macular Degeneration (wet AMD) which is the leading cause of irreversible visual loss for elder people can be subdivided into Neovascular AMD and Polypoid Choroidal Vasculopathy (PCV). Their clinical manifestations are similar, but the treatment schemes are different. To diagnose these two subtypes, Optical Coherence Tomography (OCT) is the crucial technique relied upon by ophthalmologists. Recently, the convolutional neural network (CNN) has been widely explored in the automatic diagnosis of fundus diseases including AMD. Their excellent performance proved that the CNN-based deep learning model has great potential in the fine-grained classification task of wet AMD. However, the improvement on existing methods is limited due to the tiny lesions area and the high similarity of OCT images. To overcome these challenges, in this paper, we propose a self-attention enhanced CNN framework for finegrained classification of wet AMD (SAE-wAMD) that contains the SAE-VGG16 network and a self-supervised pre-training strategy. Specifically, to capture global features and determine the location of lesions area, we designed the SAE-VGG16 model which integrates self-attention layers and convolution layers in a VGG model architecture. Furthermore, we introduced the method of contrastive learning to use self-supervised pre-training to help the SAE-VGG16 model precisely distinguish every OCT image instance and further promote its performance in the classification task. Experimental results on our OCT dataset which contains more than 4900 OCT images including neovascular AMD, PCV and normal instances show that our SAE-wAMD model outperforms state-of-the-art methods, and verify the effectiveness of our model.