A Deep Learning based Severity Grading of Diabetic Retinopathy using Data Fusion Approach

Abstract

Diabetic retinopathy is a major contributor to blindness. Traditional diabetic retinopathy detection methods typically rely on time-consuming and manual examination of retinal images. In contrast, modern developments in deep learning have revolutionized this process by enabling automated classification of retinal images with high accuracy. This research introduces a 5-stage severity grading framework for classifying diabetic retinopathy through the use of data fusion and deep learning. This study incorporates 3 openly accessible datasets for diabetic retinopathy analysis: Messidor-2, IDRiD and APTOS 2019. These datasets are first combined and then preprocessing is applied to this fused dataset. CLAHE is used to enhance the overall clarity of images. To counteract class imbalance, we employed SMOTE in the dataset. Augmentation techniques including rotation, zooming and flipping are applied to enhance the diversity of the dataset. For classification, we proposed a Convolutional Neural Network called D-Retino. Composed of four blocks, this model contains fully connected, max pooling and convolutional layers. The model's training process is fine-tuned using the Adam optimization algorithm. D-Retino demonstrated an impressive 95.24% test accuracy in distinguishing between five diabetic retinopathy grades. Moreover, results obtained by this model were benchmarked against some state-of-the-art techniques and D-Retino yielded superior performance.