Advancing Visual Perception Through VCANet-Crossover Osprey Algorithm: Integrating Visual Technologies.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, requiring efficient automated screening.
• Traditional deep learning (DL) models for DR detection face challenges with subtle lesions and computational complexity.
• Existing DL models often overlook higher-order visual processing regions, limiting their effectiveness.

Purpose of the Study:

• To develop a computationally efficient deep learning model for accurate detection of subtle lesions in diabetic retinopathy.
• To enhance DR screening by integrating multi-level visual cortex emulation and advanced optimization techniques.

Main Methods:

• Introduction of the Vision Core-Adapted Network-based Crossover Osprey Algorithm (VCANet-COP).
• Integration of Sparse Autoencoders (SAEs) for pixel-level feature extraction of vascular structures and abnormalities.
• Front-end network emulates visual cortex regions (V1, V2, V4, IT); Crossover Osprey Algorithm (COP) optimizes hyperparameters using Osprey Optimization Algorithm (OOA).

Main Results:

• VCANet-COP demonstrated superior performance across multiple DR datasets (DR-Data, STARE, IDRiD, DRIVE, RFMID).
• Achieved average metrics: 98.14% accuracy, 97.9% sensitivity, 98.08% specificity, 98.4% precision, 98.1% F1-score, 96.2% kappa.
• Reported low error rates (2.0% FPR, 2.1% FNR) and fast execution time (1.5s).

Conclusions:

• VCANet-COP offers a scalable and robust solution for automated diabetic retinopathy screening.
• The model effectively addresses limitations of traditional DL methods in detecting subtle lesions.
• Provides a valuable tool for clinical decision support in DR management.