Analysis and refinement of machine learning methods for blood vessel morphometry

Abstract

The vascular system has fundamental importance for many different types of animal tissues. Through the propagation of nutrients, oxygen, and blood cells, the vascular system helps in fighting diseases and inflammation, in the homeostasis of organisms, and in several other physiological processes. As a consequence, the characterization of blood vessels is of fundamental importance in the diagnosis of diseases and also for several state-of-the-art studies in biology. One common research topic concerns blood vessel morphometry, which involves defining properties to characterize the shape and connectivity of tissue vascularization. Given the large number of blood vessels present in cellular tissues, the systematic analysis of vascular morphometry requires the application of automated or semi-automated computational methodologies for processing images acquired using different imaging modalities. In this project, four main studies related to different aspects of the computational process necessary for the identification and characterization of blood vessels will be carried out: i) the development of techniques for interpreting and improving Convolutional Neural Networks for blood vessel segmentation; ii) the study of the influence that segmentation errors have on blood vessel morphometry; iii) the generation of realistic artificial vascular networks for training Machine Learning algorithms and iv) characterization and correction of the topology of vascularization graphs using Convolutional Graph Networks. The developed methodologies will be used for studying the development and function of the blood-brain barrier in mice. (AU)