Corneal Endothelium Evaluation in Presence of Corneal Fuchs’ Dystrophy via Convolutional Neural Networks

Specular microscopy is a non-contact technique used to image the corneal endothelium (CE). The CE is primarily made up of hexagonal cells, and an accurate assessment of its health can be performed by measuring the morphometric parameters. However, traditional methods often struggle in cases of diseases, such as cornea guttata. In this thesis, we investigated the use of deep learning methods to assess CE health without manual intervention, both in healthy and pathological subjects. We present the results of using a fully convolutional regression network to predict cell density maps from input microscopy images. In addition, we present two strategies for performing cell segmentation: classification and regression. The first approach uses a 5-layer UNet architecture to classify each pixel in an input specular microscopy image into one of three categories: the cell body, diseased region, or intercellular space. The second approach, a novel regression architecture based on UNet, aims to predict signed distance maps. We compared this approach with manual references and a comercial software. Finally, we summarize our conclusions and limitations, and outline future work. Our results show that deep-learning-based methods can be a promising tool for CE health assessment, providing a more effective and automated approach.