Diabetic Retinopathy Detection from Eye Fundus Images with Parameter Tuning for Convolutional Neural Networks

Abstract

Automated Diabetic Retinopathy (DR) detection, screening and diagnosis are critical to save vision loss of patients and assist the ophthalmologists in mass screening. DR screening aims at early treatment of the disease by detecting it before the stage progresses. Present DR analysis systems use digital fundus images for diagnosis reducing the high cost of manual computation. Researchers are continuously persisting for automated screening systems which can reduce the subjective interpretation and screening burdens for ophthalmologists. This paper proposes different Convolutional Neural Network (CNN) architectures with parameter tuning for DR classification. The proposed approach overcomes the class imbalance problem by fine tuning the network parameters. Different filter size variations are considered in the design and their altering response are analyzed at the classification output layer on a benchmark retinal image dataset. CNN model proposed in this paper provides an accuracy of 87.5% with cross entropy loss of 0.6370 with processing time of 1 minute and 23 seconds. Maximum accuracy improvement of 13% is achieved by the proposed approach over state of the art methods demonstrating its preeminence in fundus image classification.