Deep learning in early keratoconus/corneal ectasia

Abstract

Machine learning algorithms have been gaining tremendous relevance in the last decade in more or less all disciplines in medicine. In most cases supervised learning techniques are used to perform classification or graduation of diseases. Early detection of keratoconus in a screening environment is crucial as in this stage several therapeutic options such as collagen crosslinking or intracorneal ring implantation are available which could prevent further progression of the disease. In this talk we focus on new techniques of automatic identification of ectatic corneal diseases such as keratoconus in an early stage based on anterior segment optical coherence measurements with the Casia 2 tomographer. A neural network approach is implemented and fed with data of corneal front surface height and curvature, back surface height and curvature, and thickness data (in total 5 data layers) and the similarity of the measurement to normal or abnormal situation is provided. Performance data are shown for our training, validation, and test dataset.