Fusion of Different Corneal Parameters to Improve the Diagnosis of Keratoconus

Purpose: To diagnose keratoconus from healthy eyes, as well as suspected keratoconus. Methods: Certain parameters were extracted from Casia, Corvis, and Pentacam HR devices for 3 groups of healthy, with keratoconus, and suspected keratoconus. This study was performed on 340 eyes with keratoconus, 310 normal eyes, and 350 suspected keratoconus. The processing method involved the fusion of features derived from 3 devices using optimal and superior feature selection method and Genetic Algorithm (GA). Common methods of z-score normalization, minimum-maximum, Softmax, and maximum value normalization were used to normalize the parameters. The classifiers used were also Multiclass Support Vector Machine (MSVM), Multilayer Perceptron (MLP), and K-Nearest Neighbor (KNN). In the GA method, the best case was shown in MSVM classification and minimum-maximum normalization with mean accuracy and a standard deviation of 87.5±2.5%. Results: The results showed the separation of three classes of keratoconus, keratoconus suspected and healthy with two methods of GA and MSVM with mean accuracy and standard deviation of 2.5 ± 87.5%. Conclusion: This result demonstrates the superiority of the principal component analysis method for the fusion of parameters derived from the three Pentacam HR, Casia, and Corvis ST devices and the robustness of the MSVM in separating these three classes.