Multi-Head Feature Extraction for Diabetic Retinopathy Detection: A Hybrid Deep Learning and Texture Analysis Framework

Abstract

Diabetic Retinopathy (DR) remains one of the most consequential complications of diabetes mellitus, progressively impairing vision and, in advanced stages, causing irreversible blindness if not detected and treated early. Existing single-backbone deep learning approaches, while showing promising results, fall short in capturing the multi-scale, heterogeneous feature landscape inherent to retinal fundus images — particularly in distinguishing subtle inter-grade pathological differences. This paper presents MHFE-DRD, a Multi-Head Feature Extraction framework for Diabetic Retinopathy Detection that synergistically combines three pre-trained convolutional neural networks — DenseNet-121, Inception-v3, and ResNet-50 — with handcrafted Gabor filter and Local Binary Pattern (LBP) texture descriptors. A novel attention-weighted feature fusion mechanism consolidates a 5,632-dimensional joint feature vector, enabling the model to simultaneously leverage global semantic representations, multi-scale spatial structures, residual depth features, and fine-grained lesion textures. Evaluated on the APTOS 2019 and IDRiD benchmark datasets, MHFE-DRD achieves an accuracy of 97.6%, sensitivity of 96.8%, specificity of 97.2%, F1-score of 97.1%, and AUC-ROC of 98.3%, surpassing all contemporary state-of-the-art methods by a statistically significant margin. Ablation experiments further validate the complementary contribution of each feature extraction head. The framework offers a computationally efficient, clinically interpretable pipeline suitable for real-world DR screening deployment.