PyUAT: Open-source Python framework for efficient and scalable cell tracking

TL;DR

The paper tackles uncertainty in tracking microbial cells when imaging frame rates are limited. PyUAT implements uncertainty-aware tracking (UAT) via Bayesian multi-hypotheses tracking with a particle filter, generating a distribution of cell lineage trees (CLTs) by solving an ILP-based frame-to-frame assignment between frames $t$ and $t+1$. Assignment models for appearance, disappearance, migration, and division are modular, with single-cell features modeled by univariate PDFs and combined into joint likelihoods; these are evaluated using vectorized CLT walks. Tailored model combinations (e.g., FO+DD and FO+G+O+DD) improve division F1 and maintain higher LNK across increasing imaging intervals, while tensor_walks and parallel ILP enable scalable runtimes (≤2 hours for 100k+ cells) and competitive performance against non-DL LAP trackers. PyUAT thus offers a transparent, extensible platform for uncertainty-aware lineage reconstruction in microbial systems and beyond.

Abstract

Tracking individual cells in live-cell imaging provides fundamental insights, inevitable for studying causes and consequences of phenotypic heterogeneity, responses to changing environmental conditions or stressors. Microbial cell tracking, characterized by stochastic cell movements and frequent cell divisions, remains a challenging task when imaging frame rates must be limited to avoid counterfactual results. A promising way to overcome this limitation is uncertainty-aware tracking (UAT), which uses statistical models, calibrated to empirically observed cell behavior, to predict likely cell associations. We present PyUAT, an efficient and modular Python implementation of UAT for tracking microbial cells in time-lapse imaging. We demonstrate its performance on a large 2D+t data set and investigate the influence of modular biological models and imaging intervals on the tracking performance. The open-source PyUAT software is available at https://github.com/JuBiotech/PyUAT, including example notebooks for immediate use in Google Colab.

Figures (4)

• Figure 1: Schematic overview of the PyUAT approach (A--E), and tracking performance comparison (F+G). In (A+E), the size of the blue boxes indicates the probability of a lineage particle (the bigger, the more likely). The edge colors in B and D indicate that non-scored (grey) and probability-scored assignments (green -- high, yellow -- medium, red -- low probability). AP and DP denote appearance and disappearance assignments, respectively. (F) shows PyUAT derived division F1 scores for different statistical model compositions and increasing imaging intervals. (G) compares the tracking quality (see text) of PyUAT with established tracking methods, measuring the LNK score at various imaging intervals (median of five time-lapse sequences).
• Figure 2: Grid search results for growth and movement scale parameters showing TRA score results (left) and tracking execution time in seconds (right).
• Figure 3: Lineage tree example (top left) and its vectorized CLT representation (top right). The nodes are labeled with indices. The lineage vector stores the inverse links of the CLT (left), i.e., for every node the parent index is stored (or -1 if no parent is available). The bottom shows the exemplary walk matrix generation starting at nodes $3,4,5,2$ and performing for two steps. Walks can be reconstructed by traversing the columns of the walk matrix.
• Figure 4: Execution time of UAT in comparison with the existing tracking methods (left) and among the different tracking configurations (right) at various imaging intervals.