Dual vision transformer with bio-inspired optimization for explainable keratoconus classification.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Keratoconus (KCN) is a progressive corneal disease causing vision loss.
• Early detection and accurate staging are crucial for effective KCN management.
• Corneal topographic maps are essential for KCN diagnosis.

Purpose of the Study:

• To develop an enhanced deep learning framework for classifying keratoconus stages.
• To improve the accuracy and reliability of KCN diagnosis using AI.
• To provide a clinical decision support tool for ophthalmologists.

Main Methods:

• Utilized a Dual Vision Transformer (DViT) for feature extraction from corneal images.
• Employed the Electric Eel Foraging Optimizer (EEFO) for DViT hyperparameter tuning.
• Integrated LIME and SHAP for model interpretability and visualization of key corneal regions.

Main Results:

• The DViT-EEFO model achieved high performance: 99.2% accuracy, 99.3% recall, and 99.5% precision.
• Interpretability methods confirmed the model focuses on clinically significant corneal areas.
• The proposed framework outperformed existing KCN classification methods.

Conclusions:

• The DViT-EEFO framework offers superior classification performance for keratoconus.
• Enhanced model interpretability increases trust and clinical applicability.
• This AI tool shows significant potential for early keratoconus diagnosis and treatment planning.