SPF-CellTracker: Tracking Multiple Cells with Strongly-Correlated Moves Using a Spatial Particle Filter

Osamu Hirose, Shotaro Kawaguchi, Terumasa Tokunaga, Yu Toyoshima, Takayuki Teramoto, Sayuri Kuge, Takeshi Ishihara, Yuichi Iino, Ryo Yoshida

TL;DR

This paper introduces SPF-CellTracker, a novel algorithm for tracking hundreds of cells in 4D (3D space + time) microscopy images. It addresses the challenge of distinguishing similar-looking cells by leveraging the fact that their movements are often correlated, outperforming existing methods in accuracy and robustness.

What Problem Does This Paper Address?

Tracking cells in live biological samples over time (4D imaging) is crucial for understanding dynamic biological processes, like neural activity in C. elegans. However, this task is fraught with challenges:

  • Visual Similarity: Cells, especially when imaged as simple globular objects, can look very similar, making it hard to tell them apart.
  • High Cell Density: In many experiments, hundreds of cells are imaged simultaneously, increasing the chances of confusion.
  • Correlated Movement: Cells in close proximity often move together due to the organism's overall motion or coordinated biological processes. Standard tracking methods often treat cell movements independently, leading to errors.
  • Cell Division: While this paper focuses on non-dividing cells, cell division adds another layer of complexity to tracking.

Existing methods, including standard particle filters, struggle with these issues, leading to frequent "cell switching" (mistaking one cell for another) and "coalescence" (multiple tracked cells merging into one). This paper aims to improve tracking accuracy and robustness by explicitly modeling the correlated movements of cells.

Methods

The core of SPF-CellTracker is a Spatial Particle Filter (SPF), an advancement over traditional particle filters. Here's a breakdown of the key techniques:

  • Particle Filter Foundation: Like standard particle filters, SPF uses a set of "particles" (hypothetical positions) to represent the probability distribution of a cell's location. These particles are updated over time based on predicted movement and observed image data.
  • Markov Random Field (MRF) for Correlation: To capture the correlated movements of nearby cells, the authors model these relationships using a Markov Random Field. This allows the movement of one cell to influence the predicted movement of its neighbors.
  • Spatial Particle Filter (SPF) - The Novelty: Instead of relying solely on temporal information (how a cell moved in the past), SPF incorporates spatial information. It constructs a Markov Random Field tree from the initial cell positions. For each frame, it tracks the "root" cell using a standard particle filter and then sequentially tracks other cells by predicting their positions based on their parent cell in the MRF tree and their own dynamics. This "sweep" through the MRF tree allows for efficient, simultaneous tracking.
  • Proposal Distribution: A key innovation is the design of a "proposal distribution" that generates particles by considering both the cell's own dynamics and the influence of its parent cell in the MRF tree. This distribution also aims to maintain the relative positions of cells as observed in the initial frames and avoid collisions between trackers.
  • Cell Detection: Initial cell locations are detected using the DP-means algorithm, an unsupervised clustering method that automatically determines the number of clusters (cells).
  • Software Suite: The authors developed a software suite called SPF-CellTracker that includes tools for data conversion, tracking, and visualization.

Datasets

  • Synthetic Data: The authors generated synthetic 4D imaging data to simulate the characteristics of real C. elegans neural data, including correlated cell movements, globular cell shapes, and occasional cell disappearance. This allowed for controlled evaluation of the algorithm.
  • Real 4D Live-Cell Imaging Data: The primary dataset used for evaluation consists of 4D live-cell imaging data capturing the nuclei of C. elegans neurons. Specifically, they used datasets D1, D2, and D3 from "Data II" reported by Tokunaga et al. (2014). These datasets contain approximately 120 neuron nuclei per experiment.
  • Cell Detection Benchmark Data: For evaluating cell detection, they used "DATA I" which consists of distinct 3D still images of nematode neurons with manually annotated ground truth cell positions.

Key Insights and Ideas

  • Leveraging Correlated Motion is Key: The most significant insight is that the correlated movement of cells, a common characteristic in biological systems, is a powerful cue for improving tracking accuracy. By explicitly modeling this, SPF-CellTracker overcomes limitations of methods that treat cells independently.
  • Spatial Information Enhances Temporal Tracking: Traditional particle filters primarily rely on temporal predictions. SPF's innovation lies in integrating spatial relationships (via the MRF tree) into the prediction step, leading to more accurate particle generation and thus better tracking.
  • MRF Tree for Efficient Multi-Cell Tracking: Structuring the MRF as a tree allows for a sequential, efficient tracking process that avoids the computational burden of full MRF optimization or MCMC sampling, while still capturing dependencies.
  • Robustness to Tracking Failures: SPF-CellTracker demonstrates a better ability to recover from tracking failures compared to standard particle filters, which tend to propagate errors.

Key Results

  • Synthetic Data Performance: SPF-CellTracker significantly reduced tracking errors (RMSE) and tracking failures compared to the standard particle filter (PF) on synthetic datasets. For example, tracking failures were reduced from ~37-41 per nucleus to ~0.78-1.43 per nucleus.
  • Real Data Performance: SPF-CellTracker outperformed the state-of-the-art method by Tokunaga et al. (2014) in terms of tracking success rates on real C. elegans neuron data. Success rates improved from ~44-80% to ~86-96% across different datasets.
  • Computational Efficiency: The CPU time of SPF-CellTracker scales linearly with the number of targets, indicating good scalability for large datasets. For tracking 100 cells over 500 frames, the CPU time was less than 120 seconds.
  • Cell Detection: The DP-means algorithm for cell detection showed comparable true positive rates to the RPHC algorithm but a higher false discovery rate.

Limitations and Future Work

  • Limitations:
    • The current method is designed for non-dividing cells.
  • The MRF structure is restricted to a tree, which might not capture all complex spatial relationships.
  • The performance on dataset D3 (with rapid nematode movement and significant body deformation) was noticeably lower, suggesting that extreme deformations can still pose challenges.
  • Future Work:
    • Expand to dividing cells: Adapting the method to handle cell division is a key next step.
  • Broader Applicability: Extending SPF to other types of 4D imaging data, such as tracking chromosome arrangements during cell division, is a promising direction.
  • Comparison with other methods: Future work will involve comparing SPF with other tracking paradigms like detection-and-linking and contour-evolution methods.

Summary updated

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 on the basis of shape and size only, (iii) the number of imaged cells in the several-hundred range, (iv) movements of nearly-located cells are strongly correlated, and (v) cells do not divide. We developed a tracking software suite that we call SPF-CellTracker. Incorporating dependency on the cells' movements into the prediction model is the key for reducing the tracking errors: the cell switching and the coalescence of the tracked positions. We model the target cells' correlated movements 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 the nuclei of C. elegans neurons in which approximately 120 nuclei of neurons were imaged, the proposed method demonstrated improved accuracy compared to the standard particle filter and the method developed by Tokunaga et al. (2014).
BibTeX
@article{hirose2018spf,
  author = {Osamu Hirose and Shotaro Kawaguchi and Terumasa Tokunaga and Yu Toyoshima and Takayuki Teramoto and Sayuri Kuge and Takeshi Ishihara and Yuichi Iino and Ryo Yoshida},
  title = {{SPF-CellTracker: Tracking Multiple Cells with Strongly-Correlated Moves Using a Spatial Particle Filter}},
  journal = {IEEE ACM Trans. Comput. Biol. Bioinform.},
  year = {2018},
  url = {https://openreview.net/forum?id=Hf2knxJXwl},
}
Figure