Modelling, Segmenting and Analyzing in Vitro Microcirculatory Networks Using a Graph and Machine Learning Based Framework

Abstract

Microcirculatory phenomena play an important role in physiological transport processes and are increasingly investigated using microfluidic platforms. However, analysis of complex networks is often constrained by fragmented workflows, manual image processing, and single vessel analysis. In particular, cell-free layer (CFL) measurement commonly relies on individual channel studies using manual or semi-automated methods. This thesis presents an integrated analytical framework for microfluidic network analysis that combines graph theory, image segmentation with deep learning, and a unified software environment. A microfluidic chip resembling the human retina was abstracted into a directed graph, enabling explicit representation of network topology and systematic association of experimental data with individual vessels. This representation supports the application of established network analyses to characterize flow organization and connectivity. To enable automated CFL image segmentation, a fully automatic red blood cell (RBC) core segmentation method based on an Attention U-Net architecture was developed. The method displayed good performance across a range of experimental conditions (e.g hematocrit level, channel width and lighting conditions) delineating the red blood cells core accurately enabling indirect estimation of CFL thickness, with performance primarily limited by ground-truth quality. These components were integrated into an integrative analysis environment that supports visualization, data management, network analysis, and comparative studies across experimental conditions. Together, the proposed framework provides a scalable and extensible tool for the quantitative analysis of biomimetic microfluidic networks, with direct applications in microcirculatory research and the study of physiologically relevant vascular systems.