Validation of Automatic Diabetic Retinopathy Screening and Diagnosis via Deep Neural Networks on Multi-modal Retinal Fundus Image Datasets

With the global burden of diabetes mellitus and the resulting increase in diabetic retinopathy (DR) patients, there is an urgent demand for automatic DR screening and diagnosis to address the problem of limited medical resources. The aim of this study is to investigate the performance of automatic DR screening and diagnosis via deep neural networks (DNNs) on different modality fundus images. First, two retrospective fundus image datasets were collected and annotated by clinical ophthalmologists. One dataset consists of fundus fluorescein angiography (FFA) images and the other consists of color fundus images. Second, a new four-stage disease severity scale for DR diagnosis is presented to strengthen the practical application of DNNs based DR screening system in treatment method selection. Third, the deep convolutional neural network models were separately trained to diagnose DR and obtain the degree of severity. The two-stage severity classification and four-stage severity classification task were performed on both datasets. Finally, good results were obtained for the four-stage severity classification on both datasets, and the results were better than those obtained by the traditional methods. The results obtained on color images were slightly better than those obtained on FFA images. The results of the two-stage severity classification are comparable to those obtained by other state-of-the-art methods.