Purpose
Given the strong clinical similarities between normal eyes and eyes with subclinical or forme fruste keratoconus (KTC), the screening of patients for refractive surgery can sometimes be challenging. This problem could be solved by machine learning algorithms that automatically and objectively classify eyes into different groups based on a combination of parameters. This work aims to compare the accuracies of 3 of such machine learning algorithms to automatically classify corneas into normal, keratoconus and suspect.
Methods
This case-control study analyzes the basic biometric parameters provided by the Pentacam for 1379 eyes (924 subjects), of which 944 had keratoconus (671 of grade 1, 201 grade 2 and 72 grade 3 of the Krumeich scale), 85 KTC-suspect eyes, and 350 normal control eyes. Within each group, 33% of the subjects were randomly selected for the test group and the remaining were used to train the classifiers, which were Naïve Bayes (NB) with kernel density estimation, Discriminant Analysis (DA) and Support Vector Machine (SVM). The dimensionality of the data was reduced either performing Canonical Discriminant Analysis (CDA) or using a Sequential Feature Selection (SFS) algorithm with NB accuracy optimization. Finally, accuracy was estimated from the results when applying the classifier to the test group and also using a Cross-Validation (CV) method on the training set.
Results
The highest accuracy in classifying KTC 1 and normal (96.57%) was achieved by NB with SFS parameter reduction. Classifying KTC 1 to 3 and normal eyes, the highest accuracy was obtained by NB without parameter reduction (97.81%). KTC-suspect and normal eyes could be distinguished with 91.95% accuracy by DA with CDA parameter reduction. Finally, KTC-suspect, normal and KTC 1 to 3 were classified with an accuracy of 90.48% by SVM without parameter reduction (see figure 1).
Conclusions
The most suitable algorithm for KTC classification depends on the comparison at hand. Reducing the number of parameters does not impair the performance of the classification algorithm. Instead, it can even improve the accuracy in some cases. Furthermore, in the particular case of SFS, it is possible to derive the best discriminant features for KTC detection, obtaining a good accuracy with only 5 or 7 parameters.
Keywords: 574 keratoconus •
465 clinical (human) or epidemiologic studies: systems/equipment/techniques •
496 detection