Abstract
Purpose :
Keratoconus (KTC) is a progressive corneal pathology that may lead to a severe corneal deformation. Early detection of the condition is a challenge for practitioners, given the great similarity between early KTC and healthy corneas. Machine learning tools have been proposed to help identify KTC by automatically classifying corneas using on a combination of parameters. This work proposes an alternative method for KTC detection based on Artificial Neural Networks (ANNs) using only corneal elevation rather than the clinical parameters traditionally used for identifying keratoconus.
Methods :
The anterior and posterior elevation of 150 healthy corneas and 150 keratoconic corneas was measured with a Pentacam. These data were fitted to a custom-made Gaussian-Biconic model to accurately characterize the anterior and posterior elevation. This model was used to identify a limited number of geometrical features, thus allowing a robust distinction between KTC and healthy corneas by means of ANNs. A random selected subset of 50 healthy and 50 keratoconic corneas, not included as training dataset, was used to validate the model.
Results :
The error of the Gaussian-Biconic model in terms of root mean squared error (RMS) amounted to RMS << 1 mm, [0.02, 0.04] mm in the center, 0.08 mm on corneal periphery. The highest accuracy in classifying KTC and normal eyes from geometrical parameters extracted from the fitted elevation maps amounted to 94.0 % during the validation step. The Best Validation Performance (BVP) of the network was << 1 mm, and 99.0 % classification success during the training step was evaluated by the mean squared error (MSE), being MSE<BVP which confirms the reliability of the learning algorithm. However, significant correlations between the extracted geometrical features of healthy and keratoconic corneas (r ≥ 0.70; p = 0.001) were found, which may compromise accuracy.
Conclusions :
The presented methodology, based on fitted corneal elevation maps, was able to distinguish between healthy corneas and early KTC with 94.0 % accuracy during model validation. In a next step feature selection technique could be implemented to eliminate the correlations between attributes, allowing to reach even higher diagnostic accuracy.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.