Deep Learning-Based Lesion-Aware Hybrid Attention Framework for Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a progressive microvascular complication of diabetes that leads to retinal damage vision loss. However, the existing DR classification models lack lesion-specific focus and rely on single-path feature extraction that leads to suboptimal accuracy. To overcome these challenges, a novel Deep learning-based Lesion aware hybrid attention (DL-LAHA) framework is proposed for early classification DR stages using fundus images. The proposed framework integrates with an advanced Recursive Sub-image Histogram Equalization (RSIHE) technique to improve the visual clarity of the gathered images from MESSIDOR dataset. A Nested V-Net is used for segmenting the retinal regions such as hemorrhages, exudates, and with improved structural preservation and lesion boundary clarity. The Duo-Feature extraction process is performed with the combination of EfficientNet and RegNetY for extracting the relevant features from pre-processed and segmented images. Moreover, the proposed DL-LAHA framework uses an entropy-based feature selection model to ensure that the most instructive features by minimizing redundancy are retained prior to classification. The selected features are fused for classification in the fully connected layers to identify the DR stages namely Normal, Mild, Moderate, Severe, or Proliferative DR. According to the experimental analysis, the proposed DL-LAHA framework achieves the overall accuracy of 98.65% based on the gathered MESSIDOR dataset. Furthermore, the proposed framework enhances overall accuracy by 1.70%, 3.01%, 1.22%, 18.27%, and 17.85% compared to DRNet13, EfficientNet-B0, Ensemble CNN, Vision transformer, and Attention-based DL network respectively.