Abstract
Tracking many cells in time-lapse 3D image sequences is an important challenging task of bioimage informatics. Motivated by a study of brain-wide 4D imaging of neural activity in C. elegans, we present a new method of multi-cell tracking. Data types to which the method is applicable are characterized as follows: (i) cells are imaged as globular-like objects, (ii) it is difficult to distinguish cells based only on shape and size, (iii) the number of imaged cells ranges in several hundreds, (iv) moves of nearly-located cells are strongly correlated and (v) cells do not divide. We developed a tracking software suite which we call SPF-CellTracker. Incorporating dependency on cells' moves into the prediction model is the key to reducing tracking errors: cell-switching and coalescence of tracked positions. We model target cells' correlated moves as a Markov random field and we also derive a fast computation algorithm, which we call spatial particle filter. With the live-imaging data of nuclei of C. elegans neurons in which approximately 120 nuclei of neurons are imaged, we demonstrate an advantage of the proposed method over the standard particle filter and a method developed by Tokunaga et al. (2014).
Introduction
Tracking individual cells in 4D (3D + time) live-cell imaging is a cornerstone of modern biology, especially in neuroimaging. However, researchers studying the brain-wide activity of the nematode C. elegans face a "perfect storm" of technical hurdles. Neuronal nuclei appear as nearly identical globular blobs, making them indistinguishable by shape or size. Furthermore, hundreds of neurons are jammed together in tight clusters, and the worm’s body movements are sudden and rapid.
Standard tracking algorithms usually rely on temporal continuity—predicting where a cell will be based on where it just was. When the organism moves erratically, this prediction fails, leading to "cell-switching" (the tracker jumps to a neighbor) or "coalescence" (multiple trackers merge onto the same cell). SPF-CellTracker addresses this by modeling the collective, correlated movement of cells.
Methods
The authors propose the Spatial Particle Filter (SPF). While a standard particle filter tracks each cell independently over time, SPF leverages the fact that if one neuron moves, its neighbors likely move in a correlated fashion.
Key technical components include:
- Markov Random Field (MRF) on a Tree: To ensure computational efficiency, spatial relationships between cells are represented as a Minimum Spanning Tree. This allows sequential updates: tracking a "root" cell first, then using its movement to predict the movement of its "children."
- The Proposal Distribution: The SPF proposal distribution incorporates the movement of the parent cell, the relative distance between cells in the initial frame, and random noise.
- DP-means Clustering: The system automatically detects the initial number and location of cells using an extension of k-means that doesn't require pre-specifying the number of clusters.
- Rejection Sampling: To prevent trackers from overlapping, the algorithm rejects hypothesized positions that would cause collisions between neighboring cells.
Results
The authors evaluated SPF-CellTracker using synthetic datasets and real 4D imaging of C. elegans neurons (approximately 120 neurons over 500 frames). The core insight is that in systems with highly correlated motion, "where my neighbor went" is a much stronger predictor than "where I was a millisecond ago."
The results demonstrate a dramatic reduction in tracking errors. In synthetic tests, the standard particle filter averaged ~37–41 tracking failures per trial, while the SPF reduced this to ~0.7–1.4 failures. On real data, SPF achieved success rates of 86% to 96%, significantly outperforming previous state-of-the-art methods which dropped as low as 44% on high-motion datasets. Furthermore, the software is highly efficient, tracking 100 cells across 500 frames in under 120 seconds on standard hardware.
Discussion and Conclusion
SPF-CellTracker provides a robust solution for high-density, high-motion neural imaging by shifting from purely temporal tracking to a spatial-temporal model. By maintaining a "memory" of the global arrangement, the system prevents the drift common in multi-object tracking.
While currently limited by the assumption of a constant cell count (no division) and a focus on centroids rather than complex morphology, the method is highly practical for high-throughput lab use. Future work may extend SPF to dynamic structures like dividing chromosomes, where the MRF tree structure would need to update dynamically as cells duplicate and separate.
Summary updated