HybridVisionNet: An advanced hybrid deep learning framework for automated multi-class ocular disease diagnosis using fundus imaging

Şafak Kılıç,
HybridVisionNet: An advanced hybrid deep learning framework for automated multi-class ocular disease diagnosis using fundus imaging,
Ain Shams Engineering Journal,
Volume 16, Issue 10,
2025,
103594,
ISSN 2090-4479,
https://doi.org/10.1016/j.asej.2025.103594.
(https://www.sciencedirect.com/science/article/pii/S2090447925003351)
Abstract: Objectives: This study aims to develop and validate an advanced hybrid deep learning architecture for accurate automated diagnosis of multiple ocular diseases using fundus imaging. Methods/Analysis: This study introduces HybridVisionNet, a cutting-edge hybrid deep learning architecture that synergizes the multi-scale feature extraction of InceptionV3 and the dense connectivity of DenseNet121. Designed for multi-class ocular disease diagnosis, the model leverages advanced preprocessing techniques, including Contrast Limited Adaptive Histogram Equalization (CLAHE), Gaussian Blurring, and High-Pass Filtering, to enhance image clarity and diagnostic accuracy. The model was trained and validated on the ODIR-5K dataset containing 5,000 patients with eight distinct ocular conditions including Diabetic Retinopathy, Glaucoma, Cataract, Age-related Macular Degeneration (AMD), Hypertensive Retinopathy, Pathological Myopia, Other Pathologies, and Normal eyes. Findings: On the ODIR-5K dataset, HybridVisionNet achieved a groundbreaking accuracy of 99.71%, outperforming existing state-of-the-art models in precision, recall, and overall reliability. Robustness and generalization were further validated through cross-dataset evaluations on MESSIDOR-2, Kaggle DR, and EyePACS, achieving an average accuracy of 97.8%. Novelty/Improvement: The key innovation lies in the novel hybrid architecture that uniquely combines InceptionV3's multi-scale feature extraction with DenseNet121's dense connectivity, enhanced by custom attention mechanisms and feature fusion strategies. This approach significantly outperforms existing single-architecture and other hybrid models, representing a substantial advancement in automated ocular disease diagnostics. These results underscore the model's potential to revolutionize automated ocular disease diagnostics, offering a clinically viable solution for detecting conditions such as diabetic retinopathy, glaucoma, and age-related macular degeneration.
Keywords: Hybrid deep learning; Ocular disease classification; HybridVisionNet; InceptionV3-DenseNet121 fusion; Automated medical diagnosis; Fundus image analysis; Computer-aided diagnosis