ResDenseNet:Hybrid Convolutional Neural Network Model for Advanced Classification of Diabetic Retinopathy(DR) in Retinal Image Analysis

Abstract

Preventing vision loss in diabetic retinopathy (DR) requires early and precise detection. Although strong feature extraction is required and there is class imbalance in the current methods, deep learning (DL) techniques have showed promise in DR classification. With components from both the ResNeXt and DenseNet designs, a unique DL architecture for DR classification is proposed in this work. A unique DL architecture that integrates DenseNet and ResNeXt components is proposed in this work.To address unique issues in DR classification, the proposed method integrates channel-wise masking with an attention mechanism. The network is able to learn from the less frequent DR stages because the channel-wise masking reduces the influence of the majority class and the attention method concentrates the network on important features. To improve interpretability and confidence in the model's predictions, the incorporation of Explainable AI (XAI) approaches is also covered.Our findings show that the suggested approach outperforms current architectures, achieving better sensitivity for differentiating DR phases at 0.82 and accuracy at 0.87. This shows that this new method has promise for improving DR categorization, which could result in earlier diagnoses and better patient outcomes.