An Explainable Multi-Scale Attention-Based Deep Learning Framework for Trustworthy Diabetic Retinopathy Detection Using Enhanced Fundus Images

CNN Framework for Automated Diabetic Retinopathy Diagnosis

Abstract

Diabetic Retinopathy (DR) is a microvascular complication caused by prolonged hyperglycaemia that damages the retinal capillary network, putting people with diabetes at risk. Damage to these tiny vessels increases vascular permeability and results in microaneurysms, hemorrhages, and hard exudates. If left untreated, progression may lead to irreversible vision loss. Early detection and timely treatment substantially reduce the risk of blindness. Current clinical diagnosis relies on ophthalmologist assessment, which results in diagnostic delays and inter-observer variability. Inter-class similarity and intra-class variability in Fundus Images (FIs), due to heterogeneity in lesion morphology, retinal pigmentation, and imaging protocols, yields diagnostic ambiguity. Conventional Deep Learning (DL)-based frameworks act as black boxes, yielding high predictive accuracy but lacking interpretable evidence of the image features about their decisions, thereby limiting clinician trust and clinical adoption. To mitigate these limitations, we propose an explainable DL-based mechanism that classifies FIs into different DR severity grades. This study presents a heterogeneous feature-extraction method inspired by residual and inception networks, combined with attention mechanism to learn disease specific features, so that the model focuses on the most relevant patterns. Residual skip connections and multi-scale inception style branches, in combination with features refinement, helps classifying lesion severity with high precision. The proposed framework begins with segmentation of FIs to extract Regions of Interests (ROIs), followed by Channel-wise Contrast Limited Adaptive Histogram Equalization (C-CLAHE) combined with Difference of Gaussian (DoG) filtering, optimizing the color channels independently and standardizing contrast to highlight fine-grained retinal structures. Geometric minority oversampling is employed to ensure balanced representation of all severity categories during training. Proposed framework incorporates Swish activation to enhance gradient propagation and employs the class-aware focal loss formulation to mitigate class imbalance issue, emphasizing hard examples. Experimental results on the IDRiD and APTOS2019 combined test-set demonstrate that the proposed model achieves screening (healthy/diseased) and lesion-severity (mild/moderate/severe/proliferative) classification accuracies of 99.27% and 75.90%, respectively. Although, discriminating the four DR stages pose challenges like severe imbalance and inter-class differences among classes, which are better aided by proposed framework. On the DDR benchmark, the model achieves comprehensive DR-grading accuracy of 80.76% on 5-class classification task. The Heterogeneous Attention Residual-Inception Convolutional Network (HARIC-Net) outperforms several state-of-the-art (SOTA) DL-based architectures, including AlexNet, ResNet, DenseNet and Inception Net, along with recent domain specific hybrids in DR-grading task. HARIC-Net’s parameter counts of ~2 million underscores its computational efficiency and robust generalization capability. Gradient-weighted Class Activation Mapping (Grad-CAM), an Explainable Artificial Intelligence (XAI) technique, is used to visualize class-discriminative regions, and to provide interpretable predictions.