The paper presents a novel method for large-scale 3D cell tracking that leverages a hierarchy of segmentation hypotheses. The key aspects are:
Segmentation and Tracking Jointly Computed: The method computes cell segments and their tracks jointly, avoiding dependence on the accuracy of a standalone segmentation method. It leverages both spatial and temporal information.
Computational Efficiency: The approach is designed to support out-of-memory processing of multi-terabyte datasets, using strategies like parallel processing and overlapping time windows.
Flexible Segmentation Input: The method can utilize existing pre-trained models or traditional image processing algorithms for segmentation, achieving reasonable performance even without ground-truth 3D annotations.
The core of the approach is building a hierarchy of segmentation hypotheses using the watershed framework. An integer linear program (ILP) is then formulated to select disjoint segments from the hierarchy that maximize the overlap between adjacent frames, respecting biological constraints like cell division.
The method is extensively evaluated on standard benchmarks like the Cell Tracking Challenge and the Epithelial Cell Benchmark, demonstrating state-of-the-art performance in both nuclei-based and membrane-based cell tracking. It also scales to a large real-world 3D+t microscopy dataset of a developing zebrafish embryo, processing over 21.5 million cell instances.
To Another Language
from source content
arxiv.org
Deeper Inquiries