Abstract:
Keratoconus is a disorder of the eye that results in progressive thinning of the cornea. It usually occurs in the second decade of life and affects both genders and all ethnicities. The estimated prevalence in the general population is 54 per 100,000. Detecting Keratoconus is typically done using corneal tomography with different imaging systems, such as the Pentacam HR. More recently, corneal biomechanics (the corneal response to stress, and the ability of the cornea to resist deformation/distortion), has become more and more used to diagnose patients with ectatic corneal disorders such as keratoconus. However, all of these techniques rely on medical experts to manually detect keratoconus based on an inspection of the cornea tomographic images and biomechanical signals.
In this thesis, we propose to utilize machine learning to automatically detect Keratoconus based on markers extracted from tomographic and biomechanical inspections of the eye. To be able to do this, we rely on various (anonymized) datasets that are manually labelled by medical experts from the American University of Beirut Medical Center (AUBMC). Given that our datasets are limited in size, we perform 5-fold cross-validation and train various state-of-the-art machine learning techniques to automatically detect keratoconus. Our models achieved cross-validation accuracies ranging from 85\% to 100\% depending on the dataset and the classification task.
