A Dual-Phase Attention-Guided Deep Learning Framework for Imbalanced Diabetic Retinopathy Detection and Feature Localization

Diabetic Retinopathy (DR) is currently among the major causes of visual disability among diabetic patients and the right diagnosis at an early stage using automated image based systems requires. Even though deep learning models have been successful in healthcare in medical imaging, they still decline when class imbalance and lesion heterogeneity occur. As a work, this paper presents a DualPhase Attention-Guided Deep Learning (DP-AGDL) architecture that can be used to detect instances of binary DR and lesion-level informative deep reading on retinal fundus images. The proposed model combines two synergistic steps: Lesion-Aware Feature Encoding (LAFE) with a adapted residual backbone and Spatial-Channel Squeeze-and-Excitation (SCSE) attention to highlight clinically useful areas, and Flection Fusion (ARF) with focal loss and dynamically changing importance to address imbalance in datasets. Transfer learning using EfficientNet increases generalization and gradient centralization and adaptive learning rate timetable increase convergence capacity. Benchmark analysis of eyePACS and Messidor-2 illustrates that DP-AGWL outperforms current CNN and transformer-based baselines in terms of sensitivity (94.7 %), and AUC (0.982). Besides, Grad-CAM visualization affirms that the model has the interpretability necessary to be used in clinical decision-making during early DR screening.