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

Keratoconus (KCN) is a progressive degenerative corneal disorder characterized by corneal thinning and cone-shaped protrusion, leading to significant visual impairment if not detected early. Accurate staging of KCN using corneal topographic maps is critical for timely diagnosis and effective treatment planning.

This study proposes an enhanced deep learning framework for KCN stage classification based on corneal topographic images. The model employs a Dual Vision Transformer (DViT) to effectively capture both local and global spatial features. To optimize model performance, the Electric Eel Foraging Optimizer (EEFO) is utilized for tuning attention weights and hyperparameters of the DViT architecture. Additionally, model interpretability is enhanced through Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), enabling visualization of corneal regions influencing classification decisions.

Experimental evaluations conducted on a keratoconus dataset demonstrate that the proposed DViT-EEFO model outperforms existing approaches, achieving an accuracy of 99.2%, recall of 99.3%, and precision of 99.5%. Interpretability analyses confirm that the model focuses on clinically relevant corneal regions during decision-making.

The proposed DViT-EEFO framework delivers high classification performance and improved interpretability, highlighting its strong potential as a reliable clinical decision support tool for early keratoconus diagnosis and treatment planning.