Spårning av nanopartiklar med iSCAT och maskininlärning

Sammanfattning

This study explores the tracking of particles of size 100 nm from iSCAT microscopy in two and threedimensions with machine learning and algorithm-based tracking methods. The study compares theaccuracy of particle position predictions in two dimensions between the LodeSTAR method and theRadial Variance Transform algorithm (RVT), how these predictions influence the creation of partic le trajectories with MAGIK, and how the diffusivities of particles depend on the chosen detectionmethod. Furthermore, the three-dimensional LodeSTAR method is modified to train on synthesizedimages of particles in varying depths in order to gain vertical prediction capabilites. The accuracy ofthe LodeSTAR method in three dimensions is evaluated by comparing the diffusivity of predicted par ticle trajectories from experimental data with their theoretical values, as well as through the analysisof trajectories and their covariances in the vertical dimension. The results for the two-dimensionalLodeSTAR model indicate that RVT – based on covariance analysis – yields detections that indicateBrownian motion of the particles, but LodeSTAR performs better in terms of the number of correctlypredicted particles and tracking trajectories over longer time periods. Furthermore, comparisons ofdiffusivity, covariance and particle tracings for three-dimensional detections suggest that the LodeS TAR method extracts information about particle depth, but that the detections lack accuracy. It issuggested that improving the synthesized data to better capture the particle shape in a greater depthrange would yield more accurate results. In summary, the LodeSTAR method shows promising futurepotential for tracking particles in three dimensions from iSCAT microscopy footage