OcuViT: A Vision Transformer-Based Approach for Automated Diabetic Retinopathy and AMD Classification

Area of Science:

• Ophthalmology
• Computer Vision
• Machine Learning

Background:

• Early detection of diabetic retinopathy (DR) and age-related macular degeneration (AMD) is crucial for preventing vision loss.
• Manual analysis of retinal fundus images is time-consuming and requires expert interpretation.
• Existing automated methods using deep learning face challenges in computational efficiency and robustness.

Purpose of the Study:

• To propose a novel approach for automated ophthalmic diagnostics using vision transformers (ViT) and transfer learning.
• To develop and validate a model, OcuViT, for accurate classification of DR and AMD.
• To address limitations of current automated systems in terms of efficiency and robustness.

Main Methods:

• Utilized a pre-trained ViT-Base-Patch16-224 model and fine-tuned it for DR and AMD classification.
• Implemented a streamlined preprocessing pipeline to convert retinal images into PyTorch tensors and standardize them.
• Validated the OcuViT model on the APTOS dataset (DR classification) and IChallenge-AMD dataset (AMD grading).

Main Results:

• OcuViT achieved superior accuracy and robustness in five-class DR and AMD grading tasks, outperforming existing CNN- and ViT-based methods.
• The model demonstrated highly competitive performance in the binary DR classification task.
• The efficient preprocessing pipeline enhanced ViT model compatibility and performance for ophthalmic images.

Conclusions:

• OcuViT effectively leverages ViT-based transfer learning for improved precision and reliability in automated ophthalmic diagnosis.
• The proposed approach offers a significant advancement in the automated detection and classification of retinal diseases.
• This method holds promise for enhancing clinical workflows and patient outcomes in ophthalmology.