Automatic Screening of Diabetic Retinopathy Images with Convolution Neural Network Based on Caffe Framework

Abstract

Objective Diabetic retinopathy (DR) is a serious complication of eye in diabetes mellitus (DM) patients. In order to automatically screen DR, we aim to use convolutional neural network (CNN) to screen DR fundus images automatically. Methods A total of 10,551 fundus images from Kaggle fundus image dataset were collected for this experiment. Firstly, the images were preprocessed by histogram equalization and image augmentation. Then, the CNN was constructed and trained with Caffe framework. Our designed CNN models were trained by 8,626 images. Finally, the performance of the trained CNN model was validated by classifying 1,925 fundus images into DR and non-DR ones. Results The performance results indicated that the CNN achieved accuracy of 75.70% in 1,925 test fundus images. Conclusions CNN model is useful to classify the DR fundus images, thus might be applicable in further DR screening program for larger DM population.