List of Papers Browse by Subject Areas Author List
Abstract
Cell tracking in time-lapse microscopy is a fundamental problem for understanding cellular dynamics and lineage development. However, it remains highly challenging due to dynamic cellular behaviours such as shape variation, appearance drift, and complex division events. Recent deep learning-based approaches have achieved remarkable progress, yet they often rely primarily on appearance and positional cues while overlooking structural relationships among neighbouring cells, limiting their robustness under morphological changes and visual ambiguity. We propose ConstTrack, a context-aware and lineage-constrained cell tracking framework that leverages a contextual constellation signature to encode the geometric configuration of neighbouring cells, enhancing robustness under appearance drift. The framework further employs a track-pool memory to maintain both active and recently disappeared trajectories, together with lineage attributes and spatial information for stable association. A lineage-constrained delayed-decision module commits parent assignments within a sliding temporal window, enforcing biological plausibility. In addition, a spatial partitioning strategy enables memory-efficient detection on dense microscopy scenes, where detections are merged in global coordinates before identity prediction to maintain global identity consistency. Experimental results on multiple cell types and datasets, including those from the Cell Tracking Challenge and DeepCell, demonstrate that our framework achieves competitive lineage accuracy compared with recent state-of-the-art methods. The code is available at https://github.com/YiwenXu/ConstTrack.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2947_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YiwenXu/ConstTrack
Link to the Dataset(s)
Cell Tracking Challenge: https://celltrackingchallenge.net/2d-datasets/
DeepCell: https://datasets.deepcell.org/
BibTex
@InProceedings{XuYiw_ConstTrack_MICCAI2026,
author = { Xu, Yiwen AND Song, Yang AND Meijering, Erik},
title = { { ConstTrack: Constellation-Guided Cell Tracking Under Lineage Constraints } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This manuscript presents a method for tracking cells in time-lapse microscopy data. The developed approach is based on learning neighbourhood relations between cells (constellations) and delayed track assignment. The method is rather clearly presented and thoroughly validated.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Interesting cell tracking method, with two key novel components: constellation learning and delayed assignment.
- Thorough validation on two public data sets.
- Superior performance compared to state-of-the-art.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Presentation of the methodology needs more strictness.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.Methodology needs to be presented in a more strict manner, as not all the variables are introduced, some of the variables are reused, and values of some of the parameters are not reported. This hampers reproducibility of the method. Perhaps a supplementary table can offer a solution.
2.I compliment the authors for performing extensive validation using, in particular, all the available data sets from the Cell Tracking and Segmentation data repository. At the same time, in my opinion, the most convincing validation is the one performed on the two data sequences from each category that were hold out by the challenge organizers.
3.Microscopy images in Figure 1 are rather difficult to read due to small size.
4.The manuscript is rather poorly illustrated; which is more a problem of the imposed page limit.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is based on the overall quality of this submission: interesting idea that results in an algorithm that outperforms state-of-the-art methods on large set of validation data.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors of ConstTrack propose a new tracking method that can incorporate the local configuration of cells into the tracking decisions. The method is primarily based on a transformer architecture for detection and tracking and also efficiently incorporates tools for handling mitotic events. The paper is generally well-written and quantifications demonstrate that the method performs competitive to the state of the art. Moreover, the method can be trained end-to-end, which is another benefit.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Interesting approach that incorporates the local configuration of cells and their neighborhoods for the tracking decisions.
- End-to-end training of the entire tracking pipeline from detection to the tracked cells.
- Thorough validations on commonly used benchmark datasets like the cell tracking challenge.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
The Deformnable DETR is used as the object detector; however, it is not fully clear if this is based on pre-trained weights or if the weights are entirely trained in your end-to-end training approach.
-
The authors mention that as part of the auxiliary track metadata, the miss counter is stored. However, it is unclear what this miss counter actually contains and how it is being updated or defined.
-
The Cell ID decoder fuses embeddings of the last states of the track; however, here it is not clear if all of the states are fused to one embedding or if the embeddings are still kept separate for the individual time points. Please comment or explain more thoroughly.
-
For the parent assignment, there is also the class “no parent.” However, as long as there are some potential neighbors, I would guess that it’s always more likely to link one of these rather than not linking anything. Is there any threshold or other cue included that would favor the “no parent” decision over the existing neighbors in the potential parent list? Please comment on that or explain how the “no parent” class is becoming more likely.
-
The authors mentioned that the cell detections are rasterized in order to be able to use the TRA metric of the Cell Tracking Challenge. Does this imply that you are using rectangular segmentations for the tracking validation?
-
In your ablations, you show that K > 6 decreases the quality. How about using smaller K values, which means less neighbors? Is 6 already the sweet spot?
-
The authors show that removing the neighborhood constellation altogether yields poor results on one dataset; however, it is unclear how this would affect the outcome on different datasets. So it would be interesting to additionally extend the ablations to all of the other datasets for a more valid proof (e.g. as supplementary table). Generally, the ablation studies seem a bit cherry-picked and only focus on demonstrations on selected datasets.
-
Unfortunately, there is no software or repository shared or mentioned, which potentially makes it difficult or impossible to perfectly reproduce the results. Please consider sharing your algorithm upon acceptance of the paper.
Minor comments:
-
The title of the paper you’re using title case; however, in all the sub-sections you’re using sentence case. Maybe make this more consistent by deciding for one or the other spelling.
-
In Figure 1, you mentioned a grid on the input. Here, it is not clear to me where the grid is depicted in the figure. Maybe clarify or add it to the figure.
-
How is the halo overlap between different regions defined, and which explicit size do you use for your experiments? Is the detection of duplicates potentially impacted by false positives and false negatives that you might observe with your detection approach?
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Nice idea and relatively simple to incorporate into existing transformer-based tracking architectures. The paper is generally well written, but there are several small points that I think could be improved in order to further raise the quality and readability of the paper. While the quantitative results on the benchmark data sets are quite convincing (though not overwhelmingly better than the sota), the ablations seem a bit cherry-picked and therefore are not sufficiently convincing. It would thus be advisable that the authors would include more ablations on other datasets to make their claim more valid.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This work proposed a framework for achieving cell tracking with lineage constraints. To reduce tracking errors in appearance drift, the authors utilized contextual features from neighbouring cells, enforcing biological plausibility in trajectory association. Besides, halo-based spatial partitioning strategy was applied to process large dense scenes.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The authors proposed to boost the performance of cell tracking with contextual information from neighbouring detections. The appearance drift of cells in dense scenes can hopefully be solved because of the increased information window. The parent ID of each cell was therefore inferred from mixed features directly. Extensive experiments were conducted to show the superiority of their framework.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
While this work is relatively interesting regarding the methodology, a few comments are listed as follows for reference: 1.The Figure 1 does not match with the content of methodology. The features of appearance and geometry are not added directly. 2.For track-pool memory, track stack T_t^m is not referred to in the following sections. Is the framework stable with different deltas? How is the information of different “state” formulated? 3.How is the one-to-one association regularized? 4.For the quantitative results in Table 1, what is the meaning of BF-C2DL-MuSC, DIC-C2DH-HeLa… 5.For the ablation study, I don’t think the number of previous frame L is the key point. Experiments on the effectiveness of Track-pool memory, delta, composition of appearance/geometry features are considerably necessary.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The effectiveness of each modules are not validated.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank all reviewers for their comments and assessment. In the camera-ready version, we will improve clarity, reproducibility, and presentation with concise revisions under the strict page limit. We will add a public repository with code and more qualitative results.
R1-C1: Notation/parameters. We will clean notation to avoid confusing symbol and clarify variables essential to the algorithm. The main text will report key settings; complete hyperparameters and configurations will be provided in the repository.
R1-C2: Figure readability. We agree with the reviewer. We will revise Fig. 1 to improve readability, enlarge the visual components where possible.
R2-C1: Detector initialization. Deformable DETR is trained from scratch, no ImageNet- or COCO-pretrained weights are used for the ResNet-50 backbone or detector components. The detector, constellation encoder, ID decoder, and parent head are optimized jointly.
R2-C2: Miss counter. The miss counter is the number of consecutive frames without assignment. It is reset to 0 when matched, initialized to 0 for newborn tracks, and incremented when unmatched. Tracks with miss <= delta are eligible for ID prediction; older tracks remain for lineage bookkeeping.
R2-C3: Last L states. The track-pool memory keeps each track history, but historical states are not separate identity candidates. For association, the most recent L states of each eligible track are aggregated into one track token; the decoder scores one token per track plus newborn.
R2-C4: No-parent. No-parent is an explicit class in the (P+1)-way parent classifier, supervised by newborn detections that are not division children. During delayed decision, the full (P+1)-class logits, including no-parent, are accumulated. If no-parent has the maximum score, no parent edge is added.
R2-C5: CTC rasterisation. Yes, predicted boxes are rasterised as rectangular instance-labelled proxy masks only and to satisfy the official CTC input format. They are not pixel-accurate segmentations, so we report TRA and Cell-HOTA but not SEG.
R2-C6: Ablations and K value. The paper reports the most diagnostic examples. Fluo-N2DH-GOWT1 is used for constellation and K because it stresses dense local association; PhC-C2DH-U373 is used for history length because it tests temporal association. K=6 is motivated by six-neighbour topology in dense 2D cell layers and selected by our K validation.
R2-C7: Halo tiling. For dense or large-FOV sequences, halo tiling is used in training and inference to fit detector computation in GPU memory while preserving boundary context. Tile width/overlap are fixed per dataset/cell type. At inference, tile detections are mapped globally and duplicates in overlaps are merged before tracking. FP/FN remain detector errors.
R2-C8: Repository/minor issues. We will add the repository link, use consistent heading capitalization, and Fig. 1 will be revised to be more visually explicit.
R3-C1: Fig. 1 fusion. Eq. (2) fuses appearance and geometry after learned linear projections; Fig. 1 used a simplified illustration. We will show this step more explicitly.
R3-C2: Track memory/delta. Each track stores its history of detection states, only the most recent L states of eligible tracks are retrieved for ID prediction. Delta defines eligibility; older tracks are retained for lineage bookkeeping.
R3-C3: Dataset names. The dataset names follow the CTC naming convention: the prefix indicates imaging modality and dimensionality, such as BF, DIC, Fluo, or PhC, and the suffix denotes the cell type or sequence, such as HSC, MuSC, HeLa, GOWT1, U373, or PSC.
R3-C4: Ablation scope. We will revise the ablation wording to make clear that the paper reports representative ablations rather than exhaustive dataset-wise tables. The reported examples are selected to test the corresponding ambiguity: dense local association for the constellation module and temporal association for short track-history memory.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a cell tracking approach leveraging the spatial context information and temporal lineage constraint. Reviewers confirmed the experiment performances. But, a long list of comments was also provided by three reviewers. Authors need to address them in the final submission. The three reviewers gave >=4 review scores and the AC agreed with the recommendation.