Advanced retinal disease detection using RHT-Net: a hybrid deep learning approach with augmented fundus imaging.

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Retinal diseases cause significant visual impairment globally.
• Accurate and scalable automated screening is crucial for early detection.
• Current screening methods may lack efficiency and accessibility.

Purpose of the Study:

• To introduce RHT-Net, a novel hybrid deep learning model for multi-class classification of nine retinal diseases.
• To evaluate the performance of RHT-Net using color fundus images.
• To assess the model's potential for automated retinal disease screening.

Main Methods:

• Developed RHT-Net, combining CNNs for local features and Transformers for global dependencies.
• Utilized a dataset of 5,318 color fundus images from Bengali patients, augmented to 21,272 images.
• Preprocessed images (224x224 resolution, CLAHE) and split into 80% training and 20% testing sets.

Main Results:

• RHT-Net achieved high performance: 97.93% training accuracy and 96.10% F1-score.
• The model attained 96.12% accuracy and 92.28% F1-score on the test set.
• Overall classification performance reached 97.57% accuracy and 95.31% F1-score, with robust and interpretable predictions.

Conclusions:

• RHT-Net demonstrates a promising, scalable approach for early retinal disease screening.
• The model's ability to capture local and global features enhances classification performance and interpretability.
• RHT-Net has potential for integration into telemedicine and remote diagnostics, pending further validation.