Some automatic methods have been proposed to identify keratoconus from corneal maps; among these methods, neural networks have proved to be useful. However, the identification of the early cases of this ocular disease remains a problem from both a diagnostic and a screening point of view. Another problem is whether a keratoconus screening must be performed taking into account both eyes of the same subject or each eye separately; hitherto, neural networks have only been used in the second alternative. In order to examine the differences of the two screening alternatives in terms of discriminative capability, several combinations of the number of input, hidden and output nodes and of learning rates have been examined in this study. The best results have been achieved by using as input the parameters of both eyes of the same subject and as output the three categories of clinical classification (normal, keratoconus, other alterations) for each subject, a low number of neurons in the hidden layer (lower than 10) and a learning rate of 0.1. In this case a global sensitivity of 94.1% (with a keratoconus sensitivity of 100%) in the test set as well as a global specificity of 97.6% (98.6% for keratoconus alone) have been reached.
