Local–Global Feature Fusion Using CNN and Vision Transformer with Ensemble Post-Classification for Diabetic Retinopathy Diagnosis

Diabetic retinopathy is a leading cause of vision impairment globally, necessitating timely and accurate diagnosis to prevent irreversible damage.This paper proposes a novel hybrid deep learning framework that combines local and global feature representations for robust DR classification from retinal fundus images.Local features are extracted using a convolutional neural network branch that captures fine-grained pathological patterns such as microaneurysms and hemorrhages.Simultaneously, global contextual features are learned through a Vision Transformer, which models long-range dependencies across the retinal image.The extracted features from both branches are fused and passed through a series of dense layers for initial classification.To further enhance generalization and interpretability, features from the Global Average Pooling layer are used to train a Random Forest classifier.The proposed methodology is evaluated on a benchmark DR dataset with five severity classes.Extensive experiments and ablation studies demonstrate the effectiveness of our architecture in capturing both fine-grained and holistic features, leading to improved classification performance.Our results suggest that the fusion of local and global features, combined with ensemble post-classification, can provide a robust and scalable solution for automated DR diagnosis.