MaxGRNet: A multi-axis vision transformer with improved generalization for eye disease classification using...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Computer Vision

Background:

• Eye diseases like diabetic retinopathy, glaucoma, and cataracts pose significant global health risks, potentially leading to blindness.
• Timely identification of these conditions is crucial for effective treatment and vision preservation.
• Current automated methods for analyzing ophthalmic images require further enhancement in accuracy and transparency.

Purpose of the Study:

• To introduce a novel eye disease classification framework utilizing a multi-axis vision transformer (MaxViT).
• To integrate Explainable Artificial Intelligence (XAI) techniques for enhanced model transparency and interpretability.
• To evaluate the framework's performance against established deep learning models for ophthalmic image analysis.

Main Methods:

• Development of a MaxViT architecture incorporating transformer attention and GRN-based MLP layers for complex feature extraction.
• Application of the framework to a public dataset of color fundus images for eye disease classification.
• Rigorous data preprocessing and a five-fold cross-validation strategy to ensure model robustness and generalizability.
• Evaluation using metrics such as accuracy, precision, and recall, with statistical significance testing.

Main Results:

• The proposed MaxViT framework achieved superior performance compared to CNNs (ResNet50) and ViT variants (Swin-T, MaxViT-T, ViT-B16).
• Macro-averaged test accuracy, precision, and recall reached 96.75%, 96.70%, and 96.80%, respectively, with statistically significant improvements.
• XAI techniques provided valuable insights into the model's decision-making process, enhancing interpretability.

Conclusions:

• The MaxViT-based framework demonstrates robust and computationally feasible performance for automated fundus image classification.
• This approach holds significant potential for advancing decision-support systems and research in ophthalmic applications.
• The integration of advanced transformer architectures offers a promising direction for future automated eye disease diagnosis.