Abstract
Purpose: :
To compare the ability of automated pattern recognition in corneal topography with Zernike polynomial expansion among multiple regression, k-nearest-neighbor, decision-tree, and neural network.
Methods: :
For the subjects of 51 keratoconic eyes, 46 keratoconus suspect eyes, 50 myopic LASIK eyes, and 65 normal eyes, Placido-based videokeratography was performed with KR-9000PW (Topcon) for anterior corneal surface, and the corneal wavefront aberrations were analyzed with the 2nd to 4th order Zernike polynomials for 4 mm diameter. One half of the subjects was randomly chosen as the training set and the other half was used as the test set. Using the training set, the linear multiple regression, k-nearest neighbor, decision tree, and neural networks were applied for automated classification of the subjects into four categories. Effectiveness of methods was evaluated to show the accuracy in the test set.
Results: :
The accuracies of the linear multiple regression, k-nearest neighbor, decision tree, and neural networks were 80%, 70%, 69%, and 77%, respectively. Although the sensitivity for the normal control and post-LASIK groups were more than 60% for all the programs, the sensitivity of keratoconus suspect group was less than 50% because of the misclassification between keratoconus and keratoconus suspect. When a three-group classification was evaluated by combining keratoconus and keratoconus suspect into one category, the accuracies were improved from 80 % to 90 % for all the detection schemes.
Conclusions: :
Zernike polynomial expansion of wavefront aberrations with anterior corneal surface can be used to differentiate keratoconus, keratoconus suspect, post-myopic LASIK, and normal control with reasonable accuracy. Although linear multiple regression, k-nearest neighbor, decision tree, and neural networks showed similar results, optimized methods should be investigated to improve the accuracy.
Keywords: topography • keratoconus • discrimination