Thrombus formation is a complex process including platelet adhesion to matrix proteins like collagen, thrombus growth due to platelet-platelet and platelet-leukocyte interactions and finally thrombus resolution in response to fibrinolytic activity. Microfluidic devices like (coagulation) flow chambers allow to study this process over time, but bright field light microscopy provides images with an overall poor contrast. Fluorophore-labeled antibodies against cellular components (platelet surface receptors) or fibrinogen allow immunofluorescence images to complement the bright field microscopy. Nevertheless, quantification of these images is labour-intensive and strongly user-based. Machine-based learning approaches allow to elucidate the low-contrast high-content images not only to replace and standardize the image analysis, but allows also to disclose yet unknown parameters accompanying platelet adhesion, aggregation and thrombus resolution.

Project P08 aims to develop imaging pipelines to study thrombus formation under inflammatory conditions. Immunothrombosis is for example a hallmark of sepsis and usually investigated in flow chamber-based assays, however often  limited by low-contrast high-content images. Since scoring of these images is challenging, Harald Schulze and Katrin Heinze will explore machine-learning and deep-learning tools to classify and quantify such immunothrombotic events with minimized bias.

ECR.P08.11 Scoring thrombus formation from high-content low-contrast images under flow by artificial intelligence-based image analysis

– will  (i) analyze and score already acquired datasets from genetically modified mouse strains and from patients with sepsis and COVID-19 compared to healthy controls by machine-based algorithms  (ii) will further explore imaging pipelines supported by supervised and (semi)-supervised algorithms to decipher yet unknown parameters that accompany thrombus initiation, growth, stability and resolution.