Data-Driven Modelling of Corneal Diseases Using Machine Learning

Abstract

Early corneal ectatic disease remains difficult to detect because the most clinically obvious signs tend to emerge only at relatively advanced stages, limiting the effectiveness of early screening. This challenge is compounded by variability in how corneal structure is represented and interpreted, making subtle abnormalities harder to identify reliably. Improving diagnosis therefore requires not only stronger predictive models, but also more clinically meaningful representations of corneal structure. The thesis presented here addresses this problem through interpretable data-driven modelling across two related corneal ectatic conditions: keratoconus and post–laser vision correction ectasia. For keratoconus, classification was first studied using Zernike features derived from corneal pachymetry, showing that a compact and clinically interpretable thickness representation can separate normal, suspect, and keratoconic eyes with strong performance while highlighting the importance of abnormal thickness progression and asymmetry. This was complemented by the development of curvature-based indices to reassess the role of posterior curvature, demonstrating that posterior curvature is more informative than traditionally appreciated when analysed with sector-based rather than fixed-point asymmetry features. In the postoperative setting, post–laser vision correction ectasia was investigated using both engineered features and raw corneal maps, where direct learning from raw maps provided a modest but consistent improvement over handcrafted summaries.