VR-FuseNet: A Fusion of Heterogeneous Fundus Data and Explainable Deep Network for Diabetic Retinopathy Classification

Abstract

Diabetic retinopathy is a serious ocular disorder resulting from diabetes, characterized by damage to the retinal blood vessels, which may result in visual impairment and blindness if left untreated. To address the constraints of current approaches, such as dataset imbalance, diversity, and generalization difficulties, this study introduces VR-FuseNet, a novel hybrid deep learning model for automated diabetic retinopathy (DR) detection. VR-FuseNet integrates the advantages of two advanced convolutional neural networks: VGG19, which excels in capturing intricate spatial characteristics, and ResNet50V2, recognized for its profound hierarchical feature extraction capabilities. A hybrid dataset composed of five publicly accessible diabetic retinopathy datasets (APTOS 2019, DDR, IDRiD, Messidor 2, and Retino) is employed, incorporating critical preprocessing methods including the Synthetic Minority Over-sampling Technique (SMOTE) for class balancing and Contrast Limited Adaptive Histogram Equalization (CLAHE) for image enhancement to enhance robustness and generalizability. The proposed VR-FuseNet model attains exceptional diagnostic performance with > 91% accuracy, exceeding > 92% in precision, recall, and F1-score, and achieving > 98% AUC, surpassing individual constructs across all performance parameters. To improve clinical utility and interpretability, various Explainable Artificial Intelligence (XAI) techniques, including gradient-based methods, are employed to produce visual explanations that distinctly highlight retinal features influencing the model’s predictions, thereby allowing clinicians to interpret and validate diagnostic decisions.

Keywords

Citations by Year

Entered manually, not live-tracked.