Representation learning and applications in retina imaging

My Master Thesis concerns methods for the segmentation of retina fundus images based on innovative techniques from deep learning. Systematic diseases such as diabetic retinopathy, glaucoma and aged-related macular degeneration, are known to cause quantifiable changes in the morphology of the retinal microvasculature. This microvasculature is the only part of the human circulation that can be visualized non-invasively in vivo so that it can be readily photographed and processed with the tools of digital image analysis. As the treatment of serious pathologies such as diabetic retinopathy can be significantly improved with early detection, retinal image analysis has been the subject of extensive studies. To carry out this task successfully, one needs to quantify the morphological characteristics of the vascularization of the retina. However, manual extraction of this information is time-consuming, labor-intensive and requires trained personnel. For this reason, several methods have been proposed for the automated segmentation of the retinal microvasculature. Thanks to the advances in image processing and pattern recognition during the last decade, a remarkable progress is being made towards developing automated diagnostic systems for diabetic retinopathy and related conditions. Despite this progress, several challenges remain. With recent remarkable advances in the field of neural networks and deep learning, several improved methods for segmentation have been introduced in biomedical imaging. Unlike classical model-based methods, neural networks require a training stage, hence there is the need of training data, specifically images annotated by domain experts. Nonetheless, Convolutional Neural Networks (CNN) like U-net can be trained with a relatively small number of training examples. One main advantage of neural networks is that they offer the possibility to extract features from raw images avoiding the need of building hand-designed features. Their ability to discover spatial local correlations in the data at different scales and abstraction levels, allows them to learn a set of filters that are useful to correctly segment the data and, at the same time, to learn a representation of their morphological characteristics. Since ocular fundus imaging is widely used to monitor the health status of the human eye and other pathologies (e.g., diabetic retinopathy) and several annotated databases are available, my investigation was focused on the development of a CNN for the segmentation of retina fundus images. For this goal, I adapted a U-net architecture consisting of two sections of convolutional filters; the first section is an encoder that is designed to find an efficient representation of the image in terms of high-dimensional feature vectors; the second section is a decoder that map the feature vectors of the encoder into an appropriate segmentation mask. By adapting existing results from the literature, I have also investigated the inclusion of an additional layer aimed at extracting an internal representation of the network encoding the morphological characteristics of the image. The training and tuning of this CNN are the core of my Master Thesis, motivated by the goal to build an automated algorithm for the segmentation of retinal images with the ability to learn the morphology of the retinal microvasculature and to output feature vectors encoding this critical information.