Classification of Healthy and Pathological Human Corneas by the Analysis of Clinical SD-OCT Images Using Machine Learning

Photorefractive Keratectomy (PRK) is a widely used laser-assisted refractive surgical technique. While generally safe, in some cases it leads to subepithelial inflammation or fibrosis. We here present a robust, machine learning based algorithm for the detection of fibrosis based on spectral domain optical coherence tomography (SD-OCT) images recorded in vivo on standard clinical devices. The images first undergo a treatment by a previously developed algorithm for standardisation. The analysis of the pretreated images allows the extraction of quantitative parameters characterizing the transparency of human corneas. We here propose an extension of this work. Our model is based on 9 morphological quantifiers of the corneal epithelium and in particular of Bowman's layer. In a first step it is trained on SD-OCT images of corneas presenting Fuchs dystrophy, which causes similar symptoms of fibrosis. We trained a Random Forest model for the classification of corneas into "healthy" and "pathological" classes resulting in a classification accuracy (or success rate) of 97%. The transfer of this same model to images from patients who have undergone photorefractive keratectomy (PRK) surgery shows that the model output for probability of healthy classification provides a quantified indicator of corneal healing in the post-operative follow-up. The sensitivity of this probability was studied using repeatability data. We could therefore demonstrate the ability of artificial intelligence to detect sub-epithelial scars identified by clinicians as the origin of postoperative visual haze.