A Comparative Study on Deep Networks for Glaucoma Classification

Abstract

The purpose of this study is to classify glaucoma and non-glaucoma images from REFUGE dataset of fundus images. Due to the imbalance of dataset, we did data augmentation and preprocessing for dataset first (including feature extraction and enhancement). We then tested the performance of some deep convolutional neural networks as baselines, including ResNet, GoogLeNet, and VGGNet. Later we introduced self-attention layer into our CNN model and tried a method based on cup-to-disc ratio. Compared to the unprocessed dataset, the processed (data augmentation and feature enhancement) dataset gave a better performace. And self-attention model also improved performance beyond original CNN. Finally our method base on the cup-to-disc ratio was way better than the CNN models above.