Automated diabetic retinopathy classification using vision transformers on optical confocal microscopy images

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection and diagnosis of DR are crucial for effective management.
• Automated diagnostic tools can aid clinicians in identifying DR.

Purpose of the Study:

• To evaluate the efficacy of the Vision Transformer architecture for classifying diabetic retinopathy in retinal images.
• To assess the model's performance on a dataset of normal and abnormal retinal images.
• To determine the potential of the AI model for clinical application in DR screening.

Main Methods:

• Utilized the Vision Transformer deep learning architecture.
• Trained and validated the model on a dataset of retinal images categorized as 'normal' or 'abnormal'.
• Evaluated model performance using accuracy, precision, recall, and F1-scores over 30 epochs.

Main Results:

• Achieved high training (98.91%), validation (98.79%), and perfect test accuracy (100%).
• Demonstrated perfect classification performance with precision, recall, and F1-scores of 1.00 for both classes on the test set.
• Observed consistent accuracy improvements and loss reductions during the 30-epoch training process.

Conclusions:

• The Vision Transformer model shows high potential for accurate diabetic retinopathy detection in clinical settings.
• The model's performance suggests it can be a valuable tool for assisting healthcare professionals in DR diagnosis.
• Future research should focus on improving model generalizability and integrating additional clinical data.