Finding and following: a deep learning-based pipeline for tracking platelets during thrombus formation in vivo and ex vivo

Abstract

The last decade has seen increasing use of advanced imaging techniques in platelet research. However, there has been a lag in the development of image analysis methods, leaving much of the information trapped in images. Herein, we present a robust analytical pipeline for finding and following individual platelets over time in growing thrombi. Our pipeline covers four steps: detection, tracking, estimation of tracking accuracy, and quantification of platelet metrics. We detect platelets using a deep learning network for image segmentation, which we validated with proofreading by multiple experts. We then track platelets using a standard particle tracking algorithm and validate the tracks with custom image sampling — essential when following platelets within a dense thrombus. We show that our pipeline is more accurate than previously described methods. To demonstrate the utility of our analytical platform, we use it to show that in vivo thrombus formation is much faster than that ex vivo. Furthermore, platelets in vivo exhibit less passive movement in the direction of blood flow. Our tools are free and open source and written in the popular and user-friendly Python programming language. They empower researchers to accurately find and follow platelets in fluorescence microscopy experiments.

Plain Language Summary

In this paper we describe computational tools to find and follow individual platelets in blood clots recorded with fluorescence microscopy. Our tools work in a diverse range of conditions, both in living animals and in artificial flow chamber models of thrombosis. Our work uses deep learning methods to achieve excellent accuracy. We also provide tools for visualizing data and estimating error rates, so you don’t have to just trust the output. Our workflow measures platelet density, shape, and speed, which we use to demonstrate differences in the kinetics of clotting in living vessels versus a synthetic environment. The tools we wrote are open source, written in the popular Python programming language, and freely available to all. We hope they will be of use to other platelet researchers.

Keywords:

• Deep learning
• image analysis
• platelets
• segmentation
• thrombosis
• tracking

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/09537104.2024.2344512.

Additional information

Funding

This work was supported by the National Health Medical Research Council of Australia (NHMRC) via grants to J.R.H. [Project Grant 1137508 & 1187595]. N.B received funding from the Swedish Heart-Lung Foundation [2017-0318], Lions Forskningsfond and Region Östergötland. J.NI is supported by a CZI Imaging Software Fellowship grant (CZI SVCF 2018-192630, 2021-238853, 2023-332000). A.S.M. is supported by an Australian Government Research Training Program Scholarship.