Diabetic Retinopathy Detection Based on Deep Convolutional Neural Networks for Localization of Discriminative Regions

Abstract

Diabetic Retinopathy (DR) is the leading cause of avoidable vision impairment. Currently, manual DR detection is a time consuming task, which relies on well-trained clinicians with skills. In this paper, we propose a novel and automatic diabetic retinopathy (DR) detection method using deep convolutional neural networks (DCNNs). To identify the region of interests (ROIs), we design an attention mechanism for scoring the specific regions, refered as regions scoring map (RSM). The RSM is based on deep convolutional neural networks, which are trained only with image-level labels on a large scale DR dataset. Specifically, the RSM is mainly inserted into deep residual networks between intermediate stages. With RSM, the proposed model can score the different regions of an retina image to highlight the discriminative ROIs in terms of image severity level. In experiments, around 30000 colour retinal images are used to train the proposed model and around 5000 images are collected to evaluate its classification performance. The results show that our DCNN model can obtain comparable performance while achieving the merits of providing the RSM to locate the discriminative regions of the input image.