A Hybrid Deep Learning Approach With Explainable AI for Diabetic Retinopathy Classification

This project presents an advanced deep learning framework for automated Diabetic Retinopathy (DR) classification, designed to assist healthcare professionals in the early detection and diagnosis of retinal diseases from fundus images. The framework utilizes retinal image data collected from multiple diabetic retinopathy severity categories and employs comprehensive image preprocessing techniques, including image resizing, normalization, and data augmentation, to improve model robustness and generalization. The dataset is divided into training and validation subsets to ensure reliable model evaluation and performance assessment. The proposed study investigates and compares several state-of-the-art deep learning architectures, including Swin Transformer, CoAtNet, DeiT (Data-efficient Image Transformer), and a novel CNN-Transformer Hybrid Network. Transfer learning is employed using pretrained models to leverage rich visual feature representations and accelerate convergence. The CNN-Transformer Hybrid model integrates the local feature extraction capabilities of EfficientNet-B0 with the global contextual learning ability of Transformer encoders, enabling effective identification of complex retinal abnormalities associated with diabetic retinopathy. The models are trained using the AdamW optimizer and evaluated through classification metrics, confusion matrices, and validation accuracy. Experimental analysis demonstrates the effectiveness of transformer-based architectures and the proposed hybrid approach in capturing both fine-grained and global retinal features, leading to accurate classification of diabetic retinopathy stages. The trained models are saved for deployment and real-time inference, enabling automated diagnosis from unseen retinal images. The proposed framework offers a scalable, reliable, and intelligent solution for computer-aided diabetic retinopathy screening, supporting early intervention and reducing the risk of vision loss among diabetic patients.

Keywords: Diabetic Retinopathy Classification, Deep Learning, Swin Transformer, CoAtNet, DeiT, CNN-Transformer Hybrid Network, EfficientNet-B0, Transfer Learning, Medical Image Analysis, Fundus Imaging, Automated Disease Detection, Computer-Aided Diagnosis, Healthcare Analytics, Vision Transformer, Retinal Disease Screening.