Computer vision approaches for the assessment of corneal ectasia screening

Abstract

Corneal ectatic disorders, including keratoconus and its subclinical forms, pose a significant challenge in refractive surgery screening due to their progressive nature and the difficulty of early detection using conventional indices. In recent years, deep learning (DL) approaches have emerged as promising tools for automated corneal image analysis, enabling the extraction of complex spatial patterns directly from imaging data. This study presents a systematic evaluation of multiple deep learning architectures for corneal ectasia screening using Scheimpflug-based corneal maps. A publicly available dataset comprising normal, keratoconus, and suspect keratoconus cases was used to train and evaluate convolutional neural networks and transformer-based models under a unified training and validation protocol. Multiple Classification scenarios were explored, including binary (keratoconus vs. normal) and multiclass (keratoconus vs. normal vs. suspect) settings. The results demonstrate that CNN-based models achieve strong and consistent performance in binary Classification, with accuracy exceeding 91%, supporting their potential role as clinical decision support tools for refractive surgery risk assessment. However, performance declined notably in the multiclass setting, particularly for the suspect category, reflecting both the intrinsic ambiguity of intermediate disease stages and limitations associated with case-level labelling. Vision transformer architectures showed inferior performance across all scenarios, likely due to dataset size constraints and training requirements. Overall, this work highlights the strengths and current limitations of DL-based approaches for corneal ectasia screening and underscores the importance of dataset design and annotation granularity to enable robust clinical translation.