Metaheuristic-optimized swin transformer with SHAP explainability for keratoconus classification from corneal topography maps.

Abstract

Keratoconus (KCN) is an uncommon corneal disorder where the central cornea undergoes advanced thinning and causes non-uniform astigmatism. This results in metamorphopsia and potential vision loss if it is left untreated. Early detection of KCN is major to provide timely intervention and address severe visual impairment. But, conventional diagnostic tools corneal topography and slit-lamp examinations are based on expert interpretation; they are subjective and insufficient to find early KCN. To overcome these challenges, this work presents an automated and scalable deep learning (DL) model for KCN detection using corneal imaging data. The model combines Improved Swin Transformer Blocks (ISTB) with Residual Multi-Layer Perceptrons (R-MLP) for capturing local microstructural irregularities and global curvature patterns in corneal topography images. Then, the metaheuristic algorithm Polar Fox Optimizer (PFO) is presented for enhancing model convergence and robustness. Moreover, the suggested work combines SHapley additive exPlanations (SHAP) explainability for providing insights into the decision making that enhances understanding of the detection. Experimental outcomes on the benchmark dataset show that the suggested model attained high accuracy (99.4%), and outperformed other models. The approach has significant ability for deployment in clinical environments in low-resource settings, by providing understandable, real-time, and expert independent diagnosis of KCN.