henhao
Sign in
BioinformaticsbioRxiv
Preprint

Bringing TrackMate into the era of machine-learning and deep-learning.

Dmitry ErshoviD, Minh-Son PhaniD, Joanna W. PylvänäineniD, Stéphane U. RigaudiD, Laure Le BlanciD, Arthur Charles‐OrszagiD, James R. W. ConwayiD, Romain F. LaineiD, Nathan H. RoyiD, Daria BonazziiD, Guillaume DuméniliD, Guillaume JacquemetiD, Jean-Yves TinéveziD

Abstract

TrackMate is an automated tracking software used to analyze bioimages and distributed as a Fiji plugin. Here we introduce a new version of TrackMate rewritten to improve performance and usability, and integrating several popular machine and deep learning algorithms to improve versatility. We illustrate how these new components can be used to efficiently track objects from brightfield and fluorescence microscopy images across a wide range of bio-imaging experiments.

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
92
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (20 Sept 2021) · CC BY 4.0

Rewritten Fiji plugin TrackMate now uses StarDist, ilastik, Weka, MorphoLibJ or imported label images as detectors, stores 2D object contours, and tracks shape and intensity over time.

TrackMate's earlier detectors relied only on a Laplacian of Gaussian (LoG) filter, which suits sub-resolved particles and blob-like objects but does poorly on textured or complex-shaped objects and non-fluorescence modalities, and records position but not shape. This version adds machine-learning and deep-learning segmentation inside the Fiji interface, whereas the paper notes that many deep-learning tracking approaches are problem-specific or need coding skills and specialised hardware.

Key results

  1. ShownA new application programming interface (API) lets Java segmentation tools (ilastik, Weka Trainable-Segmentation, StarDist, MorphoLibJ) act as TrackMate detectors, while results from non-Java tools can be imported as label images, binary masks or probability maps.

    Described and illustrated in Fig. 1. The rewritten data model keeps 2D object contours, so morphological features can be measured per object over time and related to motility and intensity. Detectors run on 2D and 3D images where possible. · Abstract; Figure legends

  2. SuggestedThe authors report that the machine-learning and deep-learning detectors track more accurately than the LoG detector when objects are not simple round spots.

    The comparison is in Supplementary Note 1 and Supplementary Figures 1-3. The main text gives no metric, values or number of datasets for it. · Abstract

  3. ShownCustom and built-in StarDist models automatically tracked SiR-DNA-labelled MCF10DCIS.com nuclei during collective migration, label-free activated T cells in brightfield, and MDA-MB-231 nuclei whose ERK reporter signal was read per cell in a second channel.

    Fig. 2a (spinning-disk confocal, Movie 1); Fig. 2b (T cells on ICAM-1, cells coloured by mean track speed, Movie 2); Fig. 2c (ERKKTRClover reporter, wide-field imaging over 17 hours, ERK activity and instant velocity shown as heatmaps, Supplementary Figure 4, Movie 3). The demonstrations are qualitative; no tracking accuracy is reported. · Abstract; Figure legends

  4. ShownPixel classifiers and imported segmentations extend tracking beyond nuclei: ilastik segmented growing Neisseria meningitidis to give lineage trees with per-bacterium area and circularity, a Weka classifier enabled focal-adhesion tracking, and cellpose label images of mouse hematopoietic stem cells were tracked.

    Fig. 2d, Movie 5: area falls sharply at each division then rises almost linearly, and circularity stays roughly flat during growth before dropping ahead of division. Focal adhesions in endothelial cells: Supplementary Figure 5, Movie 4. Stem cells in a hydrogel microwell, segmented with cellpose in ZeroCostDL4Mic: Fig. 2e, Supplementary Figure 6, Movie 6. · Abstract; Figure legends

  5. ShownLinking per-plane StarDist detections across successive z planes, rather than time points, lets TrackMate assemble 3D nuclei. This is shown on DAPI-stained MCF10DCIS.com spheroids.

    Fig. 2f, Supplementary Figure 7, Movies 8-10: nuclei were detected in each z plane and linked within TrackMate, then exported as a 3D label image. The authors suggest this speeds up 3D segmentation with models trained on 2D images; no accuracy or timing is reported. · Abstract; Figure legends

Methods, data and reproducibility

The Fiji tracking plugin TrackMate was rewritten with a new detector API and a data model that stores 2D object contours. Java segmentation tools were wrapped as detectors, external segmentations were made importable, and the result was applied to brightfield, fluorescence time-lapse and 3D datasets.

  • Time-lapse of collectively migrating MCF10DCIS.com breast cancer cells, nuclei labelled with SiR-DNA, spinning-disk confocal (Fig. 2a) · human (inferred from the MCF10DCIS.com cell-line name; species not stated)
  • Brightfield time-lapse of activated T cells migrating on ICAM-1 (Fig. 2b)
  • MDA-MB-231 cells expressing the ERKKTRClover reporter and labelled with SiR-DNA, wide-field fluorescence over 17 hours (Fig. 2c) · human (inferred from the MDA-MB-231 cell-line name; species not stated)
  • Focal adhesions in endothelial cells (Supplementary Figure 5)
  • Growth of Neisseria meningitidis expressing PilQ-mCherry, spinning-disk confocal (Fig. 2d) · Neisseria meningitidis
  • Mouse hematopoietic stem cells migrating in a hydrogel microwell, brightfield, segmented with cellpose (Fig. 2e) · mouse
  • DAPI-stained MCF10DCIS.com 3D spheroids imaged by spinning-disk confocal (Fig. 2f) · human (inferred from the MCF10DCIS.com cell-line name; species not stated)
  • Test datasets released in a dedicated Zenodo collection · n = 11 test datasets
Methods & tools
TrackMate, Fiji, StarDist, ilastik, Weka Trainable-Segmentation, MorphoLibJ, cellpose, ZeroCostDL4Mic, Laplacian of Gaussian (LoG) detector
Code
https://imagej.net/plugins/trackmate/https://imagej.net/plugins/trackmate/trackmate-v7-detectors
Software & environment
Delivered as a Fiji plugin and installed by updating Fiji; no version numbers or dependency details are stated.
Reproducibility gaps
The URLs are ImageJ wiki documentation pages, not code; the software ships through the Fiji updater and no source repository URL is printed. The 11 test datasets are on Zenodo per the text, but no link or DOI is given. The LoG comparison is in supplementary material.
Limitations (3)
  • Object contours can be displayed and analysed only in 2D (the detectors themselves work on 3D images where possible); 3D object shapes are recovered by a slice-by-slice workaround that links 2D detections across z planes. · stated by the authors
  • Segmentation tools not written in Java, such as cellpose, do not run inside TrackMate; their outputs must be produced separately and imported as masks, label images or probability maps. · noted in this summary
  • The only accuracy comparison (new detectors vs LoG) is in Supplementary Note 1 and Supplementary Figures 1-3, with no tracking metrics in the main text; the showcase datasets are shown qualitatively (figures, movies), with no head-to-head comparison against other tracking software. · noted in this summary

Who should read:Read this if you track cells, nuclei, bacteria or adhesions in Fiji and want StarDist, ilastik, Weka or cellpose segmentations to feed an interactive workflow for tracking and per-object morphology.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISAnselmet '26Wu '26Fazeli '20Gomez-de-Ma '21Thomas '20Sridhar '18
64% semantic matchMicrobiology
Marie Anselmet, Laura Xénard, Marvin Albert+15 more·Current Protocols·May 1, 2026