Recent Applications of Deep Learning for Retinal Image Analysis

Recent Applications of Deep Learning for Retinal Image Analysis

Affiliation: Postdoctoral Researcher at INESC TEC Porto

Place: Large Lecture Room

Abstract:

The human eye is an excellent source of biomarkers that are useful for the early detection of different diseases like diabetes or hypertension. Acquisition of images of the retinal fundus is becoming more affordable, and mass screening programs are routinely implemented all around the world. Automatic Retinal Image Analysis is a key tool for the efficient analysis of the large quantity of data generated by these programs, potentially leading to a better vision healthcare. In this talk we present three recent applications on retinal image analysis that rely on deep neural networks in different ways. First, we introduce a new system for the classification of retinal vessels into arteries or veins. This is a challenging pixel-wise classification problem, with relevant clinical applications, e.g. an abnormal arteriolar-venular ratio has been correlated with hypertension and coronary heart diseases. Second, we give an overview of a recently presented retinal image synthesis model, that is trained in an adversarial fashion to create retinal images out of associated binary vessel maps. We will also present an extension of this system that avoids the dependency on these vessel maps, allowing to synthesize new images by simply sampling a multi-dimensional Gaussian distribution. Finally, we introduce a simple application of deep neural networks for regressing the quality of a vessel tree image, resulting in the first existing no-reference retinal vessel quality metric.