Robust InceptionV3 with Novel EYENET Weights for Di-EYENET Ocular Surface Imaging Dataset: Integrating Chain Foraging and Cyclone Aging Techniques

Abstract

Abstract Predicting diabetic types from ocular surface eye images is a challenging task due to subtle variations in features and the potential overlap in presentations among different diabetic types. While AI-based algorithms have shown promise in distinguishing these nuances, gaps remain in the accuracy and adaptability of existing models, especially in the context of medical imaging for diabetes classification. This study addresses these gaps by proposing a novel integration of AI and medical imaging through the Manta-ray Foraging Optimization (MRFO) algorithm, which leverages cyclone aging (CA) and chain foraging (CF) strategies. We couple MRFO with hierarchical feature learning to optimize the InceptionV3 model, achieving optimal hyperparameter configuration and enhancing both accuracy and computational efficiency. The novelty of this work lies in the introduction of a newly curated dataset, Di-EYENET, which is specifically designed for diabetic eye studies and contains multiclass categories (Type-1, Type-2, and non-diabetic). Di-EYENET fills a significant gap in diabetic ocular research by offering a reliable, validated resource for training models on eye image datasets with distinct characteristics. Our results demonstrate that the InceptionV3 model, fine-tuned with the newly developed EYENET weights, outperforms both traditional ImageNet weights and other pretrained models, showing a 2% accuracy improvement. This research highlights the potential of nature-inspired optimization algorithms and tailored datasets to enhance AI model robustness and adaptability in the context of medical disease diagnosis, particularly in the field of diabetic eye disease.

Keywords

• Ocular surface eye images
• Chain foraging
• Cyclone aging
• Metaheuristic
• Optimization