Pentacam multimodal topography-based radiomics for early keratoconus screening and staging

Early keratoconus (KC) is characterized by insidious progression, which significantly hinders clinical screening. This study introduces a radiomics analysis framework based on Pentacam multimodal topographic maps and evaluates its efficacy in identifying early KC and assessing disease progression without incurring additional examination costs. This retrospective study included 415 eyes (177 normal, 84 early KC, and 154 KC). Radiomic features were extracted from four topographic map modalities and processed via a multi-step feature selection pipeline. Classification models were constructed using the logistic regression (LR), random forest, support vector machine, and XGBoost architectures, with their performance evaluated on an independent test set. The models demonstrated excellent performance in distinguishing normal eyes from KC, achieving AUC > 0.98 across all modalities. In the screening task (normal vs. early KC), the LR model based on posterior elevation map features achieved the best performance (AUC = 0.980, 95% CI: 0.943–1.000). In the disease progression assessment task (early KC vs. KC), the model performance demonstrated a clear modality dependence. The curvature and elevation map-based models significantly outperformed the pachymetry map-based model (AUC > 0.960 vs. AUC = 0.862). A SHapley Additive exPlanations analysis revealed that the early screening model was predominantly driven by high-frequency micro-texture features, whereas the disease staging models relied on low-frequency global statistical features. The proposed method can improve the detection rate of early KC, providing potential imaging biomarkers for disease progression assessment, as well as evidence supporting the pathological hypothesis that posterior corneal changes precede stromal thickness alterations during KC development.