Keeping On Track: Deep Learning-enabled Tracking Of Platelets In Growing Thrombi
In vivo time-lapse microscopy allows us to capture vast amounts of data that, with the aid of computational methods can allow us to delineate the mechanisms underpinning biological processes. In order to extract high quality quantitative data from these sorts of images, we need to develop accurate methods to finding and tracking objects of interest. In the present research, I have developed a pipeline for analysing videos of platelets in growing blood clots. This pipeline has two main components the first is known as instance segmentation, which involves finding the pixels belonging to each platelet. For this, a type of neural network known as a u-net [1] is used to predict image features that can be post-processed using a modified watershed algorithm [2]. By comparing segmentations with human-annotated images, the deep learning segmentation was found to have good accuracy and outperform classical (not learning-based) methods in segmentation quality metrics including Variation of Information [3], Average Precision [4], and cell count difference. The second stage of the analysis pipeline involved tracking the now located platelets using the TrackPy Python package. Tracking was validated using a novel approach, by which segments of track are randomly sampled and annotated in order to estimate error rates. This pipeline was applied to analyses hundreds of clots using MASSIVE’s M3 supercomputer and highlights the utility of bioimage informatics in modern biology.