An Interpretable and Reliable Hybrid Learning Framework for Diabetic Retinopathy Grading

This repository contains the reproducibility materials associated with the research article “An Interpretable and Reliable Hybrid Learning Framework for Diabetic Retinopathy Grading.” The study presents EffTNet, a hybrid deep learning framework combining EfficientNet-B3a and Vision Transformer (ViT-B16) with Attention-Augmented Feature Fusion (AAFF) and Pair-Aware Contrastive Learning (PACL) for five-class diabetic retinopathy severity grading from retinal fundus images.The development data comprise APTOS 2019, EyePACS, and Messidor-2, while DDR and DiaRetDB1 are used for external validation. The development dataset contains 40,527 fundus images distributed across five severity classes: No DR, Mild, Moderate, Severe, and Proliferative DR.The repository provides the EffTNet implementation, original research notebook, model configuration, class mapping, dataset manifest, dataset distribution, preprocessing and augmentation protocol, data-partitioning protocol, PACL description, architecture details, evaluation procedures, and reported experimental results. These materials are provided to support transparency, methodological understanding, and reproducibility of the reported study.The original retinal fundus images are not redistributed in this repository because they are third-party datasets subject to their respective access, licensing, and usage conditions. Dataset source information and the corresponding dataset organization used in this study are provided in the repository materials.