Abstract
Purpose: :
To evaluate the ability of wavefront aberrations of the anterior and posterior corneal surface in order to classify eyes with keratoconus (KC), pellucid marginal degeneration (PMD) and normal eyes.
Methods: :
This retrospective study included 55 eyes of 32 patients with KC (group 1), 60 eyes of 43 patients with PMD (group 2) and 64 eyes of 32 subjects with normal eyes (group 3). From axial-keratometric data of the anterior corneal surface and elevation data (Orbscan IIz, Bausch & Lomb) of the posterior corneal surface a Zernike decomposition (2nd - 7th order, 6 mm pupil diameter) was performed. Explorative analysis and supervised support vector machine (SVM) classification assessed anterior and posterior corneal wavefront aberrations in their ability to divide patient eyes into groups. Cross-validation was used to estimate the classification error on unseen data. The most significant Zernike coefficients (ZC) for classification between the three groups were ascertained by SVM feature selection and sparse principal component analysis. Discriminative ability of individual ZC was assessed by receiver operating characteristic curve (ROC) analysis.
Results: :
Hierarchical clustering of anterior and posterior corneal wavefront data showed a clear grouping of patients into the three major disease groups (KC, PMD, normal) as a result of the first two hierarchical splits. The first split separated healthy from diseased eyes, the second one KC from PMD eyes. The most influential ZC to distinguish between KC, PMD and normal eyes were C22, C3-1, C2-2 and C3-3. Anterior primary vertical coma C3-1 had the highest ability to discriminate between both PMD and normal eyes (standardized area under the ROC curve [AzROC] 0.991) as well as KC and normal eyes (AzROC 0.986). Anterior primary astigmatism C22 was the most powerful ZC to discriminate between KC and PMD (AzROC 0.891). Discriminative ability of anterior wavefront aberrations was higher (average AzROC 0.91) than of posterior (average AzROC 0.78). SVM classification of all 33 anterior ZC achieved a correct classification rate of 87% for unseen data.
Conclusions: :
Our data suggest that ZC from the anterior and posterior corneal surface classify between KC, PMD and normal eyes with high accuracy.
Keywords: keratoconus • cornea: clinical science • aberrations