Particle tracking in quantitative microscopy: a modular tutorial on classical and deep learning methods

Abstract

Particle tracking is a central tool in digital microscopy for quantifying micro- and nanoscale dynamics. Selecting and validating appropriate methods remains challenging, particularly under low signal-to-noise ratios or high particle densities and for researchers with limited computational expertise. In this talk, we present a modular, Python notebook-based tutorial that organizes tracking into two stages—object detection and trajectory linking—providing a structured basis for systematic comparison of classical and deep learning approaches and stage-specific performance diagnosis. For detection, we compare intensity thresholding and Gaussian feature localization (Crocker-Grier) with convolutional neural networks, including supervised segmentation (U-Net) and a self- supervised geometric deep learning model (LodeSTAR). For linking, we evaluate nearest-neighbor and Hungarian assignment methods alongside a graph neural network classifier (MAGIK). Hands-on examples using simulated datasets with ground truth support quantitative benchmarking that identifies performance regimes and computational trade-offs. The resulting workflows are ready-to-use and adaptable across diverse scenarios.