Texture-based Analysis of Corneal Dynamics for the Detection of Forme Fruste Keratoconus.

PURPOSE: To propose CVS-omics, a radiomics framework using Corvis ST (CVS) (Oculus Optikgeräte GmbH) imaging and machine learning, for precise identification of forme fruste keratoconus (FFKC), a subtle corneal condition often undetected by conventional diagnostics.

METHODS

A total of 410 eyes were evaluated, including 265 normal eyes and 145 FFKC eyes. Texture features through radiomics were extracted from CVS images acquired at three key deformation phases: initial state, first applanation, and maximum deformation. These features were used to train three machine learning models (Random Forest [Minitab, Inc], C5.0 [RuleQuest Research], and XGBoost (extreme Gradient Boosting) on 328 eyes, with testing conducted on 82 eyes. Diagnostic performance was evaluated using receiver operating characteristic (ROC) analysis and compared with conventional biomechanical parameters.

RESULTS

The CVS-omics Random Forest model achieved superior diagnostic performance (area under the curve (AUC) = 0.989, sensitivity = 0.931, specificity = 0.962, accuracy = 0.951), significantly outperforming traditional CVS parameters (best AUC = 0.764). Models trained on features from three deformation phases showed higher diagnostic accuracy than those based on a single phase. Other models demonstrated high generalizability of dynamic radiomics features (XGBoost and C5.0 AUC > 0.83).

CONCLUSIONS

CVS-omics effectively detects subtle characteristic alterations in FFKC eyes with superior accuracy compared to conventional biomechanical parameters. This texture feature approach shows promise as a non-invasive clinical tool for FFKC detection and timely intervention.