Composition Counts: A Machine Learning View on Immunothrombosis Using Quantitative Phase Imaging

Abstract

Thrombotic complications are a leading cause of death worldwide, often triggered by inflammatory conditions such as sepsis and COVID-19, due to a close relationship between inflammation and hemostasis known as immunothrombosis. Platelet activation and leukocyte-platelet aggregation play key roles in microthrombotic events, yet there are no routine diagnostic predictive biomarkers based on these factors. This work presents a novel processing pipeline using label-free Quantitative Phase Imaging (QPI) for the detection and quantitative analysis of blood cell aggregates without sample preparation. For evaluation, we use different test scenarios and measure performance at different stages of the pipeline to gain a better understanding of the critical points. We show that, among other classical and machine learning techniques, the Mask R-CNN approach achieves the best results for detection, segmentation, and classification of cell aggregates. The method successfully identifies aggregate levels in whole blood samples and shows elevated levels in >90% of patients with COVID-19 or sepsis compared to healthy reference samples, indicating the potential of platelet and leukocyte-platelet aggregates as biomarkers for thrombotic diseases.