ViT-HHO: Optimized vision transformer for diabetic retinopathy detection using Harris Hawk optimization.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss worldwide.
• Early and accurate DR detection is crucial for preventing severe visual impairment.
• Current automated detection methods require further optimization for clinical reliability.

Purpose of the Study:

• To develop an optimized Vision Transformer (ViT) model integrated with Harris Hawk Optimization (HHO) for enhanced automated detection of diabetic retinopathy (DR).
• To evaluate the performance and generalization capabilities of the proposed ViT-HHO model on benchmark datasets.

Main Methods:

• Utilized a Vision Transformer (ViT) architecture with self-attention mechanisms for feature extraction from retinal images.
• Employed Harris Hawk Optimization (HHO) to fine-tune key hyperparameters of the ViT model.
• Validated the ViT-HHO model on the APTOS-2019 and IDRiD datasets for performance assessment.

Main Results:

• The ViT-HHO model achieved high accuracy (99.83%), sensitivity (99.78%), specificity (99.85%), and AUC-ROC (99.80%) on the APTOS-2019 dataset.
• Demonstrated strong generalization on the IDRiD dataset with 99.11% accuracy and 99.12% AUC-ROC.
• Outperformed traditional Convolutional Neural Networks (CNNs) and other optimization techniques in DR detection.

Conclusions:

• The optimized ViT-HHO model offers a highly precise and reliable approach for the automated clinical detection of diabetic retinopathy (DR).
• The model's superior performance and generalization capabilities highlight its potential to significantly enhance DR screening protocols.
• This study underscores the efficacy of combining advanced AI architectures with metaheuristic optimization for medical image analysis.