Pith. sign in

REVIEW 3 major objections 6 minor 63 references

Candidate cell links, not cells, are the right tokens for lineage tracking: geometry-biased edge attention beats node embeddings and graph topology.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An edge-centric Transformer with line-to-line geometric attention achieves SOTA cell lineage tracking without pretrained image encoders and fine-tunes far more efficiently than node-embedding baselines.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid edge-centric cell tracker that actually measures why node/GNN methods fail and then beats them on CTC and a cheap correction loop. the 3 major comments →

arxiv 2607.11754 v1 pith:NSPXAA4D submitted 2026-07-13 cs.CV

Higher-Order Cell Tracking Transformer

Jord\~ao Bragantini , Ilan Theodoro , Lo\"ic A. Royer This is my paper
classification cs.CV
keywords cell trackingedge-centric transformerline-to-line distanceparental softmaxcandidate graph heterophilylineage reconstructionhuman-in-the-loop fine-tuningCell Tracking Challenge
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Cell tracking from live microscopy is usually cast as linking detections across frames, including through divisions. The paper argues that working with cell nodes is structurally wrong: divisions force parent embeddings to sit near both daughters and therefore glue together lineage paths that should stay separate, while candidate graphs are non-homophilic (edges that share a node almost never share the same true/false label), so graph neural networks have nothing useful to aggregate. HOCT therefore treats every candidate link as its own token. After a short node stage, an edge Transformer lets those links attend to one another under a 3D line-to-line distance bias rather than under the original graph adjacency. The same edge probabilities feed a multi-gap parental softmax and a two-pass ILP that recovers valid lineages. On the Cell Tracking Challenge hidden tests a single model ranks first in overall generalizability without any deep pretrained image encoder; on a dense bacteria division set a logistic head on frozen HOCT features cuts errors by 59 percent with only 400 annotations, far ahead of LoRA fine-tuning of a node baseline.

Core claim

The right representation for cell lineage reconstruction is the candidate edge itself. Once each possible parent-to-daughter link is an independent token that attends under inter-link geometry rather than under candidate-graph topology, both the connected-manifold problem at divisions and the near-zero label agreement among co-incident edges disappear, and state-of-the-art tracking becomes possible from simple geometric features alone.

What carries the argument

Higher-Order Cell Tracking Transformer (HOCT): edge tokens built by concatenating contextualized node features, refined by self-attention whose logits are biased by the analytic 3D line-to-line distance between edge segments (attractive and repulsive heads), then normalized by a multi-frame parental softmax and decoded by a two-pass tracklet ILP.

Load-bearing premise

Every true association is assumed to already sit inside a spatially thresholded candidate graph built from given segmentations; the model can only re-weight existing candidates, not invent missing detections or links that fall outside that construction.

What would settle it

On a Cell Tracking Challenge or bacteria sequence, deliberately drop true links beyond the spatial threshold (or remove detections at divisions) and check whether HOCT's CLB/AOGM collapses relative to methods that can re-detect or search more freely; if performance holds, the candidate-graph premise is not load-bearing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Edge-centric attention with geometric biases can replace node-similarity association for any tracking domain whose candidate graphs are non-homophilic.
  • A frozen edge-feature backbone plus a cheap logistic head is enough for rapid human-in-the-loop correction, reducing the need for full LoRA or end-to-end fine-tuning.
  • Variable appearance costs derived from parental-softmax residuals give ILP solvers a data-driven alternative to fixed appearance penalties.
  • Simple hand-crafted geometric features can match or beat large pretrained visual encoders once the association architecture itself is correctly structured.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same line-to-line bias may help multi-object tracking of pedestrians or vehicles whenever identity swaps occur at near-collisions rather than at true interactions.
  • Because the method already works from 19-dimensional mask features, it is a natural drop-in for any existing segmentation pipeline that currently relies on hand-tuned association costs.
  • The O(|E|²) cost, even after k-NN sparsification, suggests hybrid designs that keep edge attention only near divisions or crowded regions while using cheaper node matching elsewhere.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes HOCT, a two-stage edge-centric Transformer for tracking-by-detection cell lineage reconstruction. Node features are refined with 3D RoPE self-attention; candidate links become tokens that attend under RoPE plus a learnable line-to-line distance bias (attractive/repulsive heads). A multi-frame parental softmax and a two-pass tracklet ILP with variable appearance costs produce the final tracks. Motivation is quantified via near-zero adjusted homophily on the candidate line graph. On CTC hidden tests a single joint model ranks first by official generalizability (CLB 0.930 / LNK 0.985 / BIO 0.875) without deep image encoders; on a bacteria-division benchmark it matches or beats Trackastra variants that use SAM2.1/DinoV2. A frozen-backbone logistic head on edge embeddings reduces AOGM by up to 59% with 400 annotations, versus at most 6.75% for LoRA fine-tuning of a Trackastra/Attrackt baseline.

Significance. If the results hold, this is a clear methodological advance for cell tracking: it identifies two structural failures of node-embedding and topology-driven GNN trackers (division-induced connected manifolds; non-homophilic candidate graphs with Hadj≈0), and shows that edge-centric attention with geometric inter-link bias addresses them. Competitive CTC generalizability and bacteria AOGM without foundation-model image encoders is practically valuable. The quantified heterophily analysis, ILP and aggregator ablations (Tables 1–2), organizer-provided hidden-test scores, and detailed appendices (architecture, ILP, correction protocol) strengthen credibility. The human-in-the-loop result, if interpreted carefully, also suggests that explicit edge representations are more annotation-efficient than node-similarity models.

major comments (3)
  1. Sec. 4.5 / Table 5 / Fig. 2: The headline fine-tuning claim (59% AOGM reduction vs 6.75% for LoRA+unsup) is load-bearing in the abstract but rests on uneven protocols. HOCT starts at AOGM 1418 vs Trackastra 4951, annotates edges rather than nodes, and only refits a logistic head on frozen features, while the strongest baseline updates LoRA adapters for 2000 steps/round after extra autoencoder pretraining on the validation sequences. Relative curves help, but the comparative claim that “edge-level representation quality dominates more complex approaches” needs either (i) a matched starting checkpoint / annotation budget in edge-equivalent units, or (ii) softer wording that emphasizes HOCT’s absolute correction efficiency rather than a direct superiority percentage.
  2. Table 2 and Sec. 3.2: Edge-Transformer (global edge self-attention without RoPE or line-to-line bias) reaches CLB 0.922±0.002 vs HOCT 0.926±0.007—within or near the reported variability—while GAT/FAGCN on L(GC) lag more clearly. The abstract and Sec. 1 attribute resolution of both structural obstacles to “attention … under a 3D geometric prior.” The ablation better supports that the edge-centric design (vs topology-restricted aggregation) is the main fix; the geometric bias is a smaller incremental gain. Please rebalance the claim so that “higher-order / edge tokens” and “line-to-line bias” are not conflated, and report whether the HOCT–Edge-Transformer gap is significant across seeds or only the two CV splits.
  3. Table 3 / Appendix E: Official CTC ranks and scores were “provided … directly” by organizers because the public leaderboard was not refreshed as of 2026-05-01. For a SOTA generalizability claim this is acceptable only with a clear verification path (e.g., organizer confirmation letter, submission ID, or commitment to update once the board is public). Please state how readers can independently verify the hidden-test numbers after publication.
minor comments (6)
  1. Fig. 1 caption and panel (e): “edge lines” / “dline” notation is dense; a short callout that d_line is the segment–segment distance of Eq. (2) would help non-specialists.
  2. Eq. (3): softplus(α_h) with α_h init −5 is sensible; state units of d_line after per-dataset standardization so the scale of the bias is interpretable.
  3. Sec. 4.1: “16 datasets … 32 sequences” and later “28 non-degenerate CTC sequences” for heterophily—briefly list which four were excluded as trivial so the Hadj statistic is reproducible.
  4. Table 4: “Original (greedy/ILP) [16]” rows lack ±std while other Trackastra rows have them; note whether those are single-run numbers from the source paper.
  5. Limitations: O(|E|²) and k-NN mitigation are clear; a rough peak-memory or wall-clock comparison vs Trackastra on one dense CTC sequence would make the cost trade-off concrete.
  6. Typos / polish: “dubbed this approachHigher-Order” (missing space); “Parental Softmax + Concat Scaling h” in Fig. 1 is hard to parse; checklist says Appendix D LoRA ~35 min while body text says ~40–55 min—align the numbers.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture paper evaluated on external public benchmarks with independent ablations.

full rationale

HOCT is an empirical ML methods paper. Its central claims (SOTA CTC generalizability CLB/LNK/BIO on hidden tests; best bacteria AOGM with hand-crafted features; 59% AOGM reduction via frozen-edge logistic head vs LoRA) are measured against external public benchmarks (CTC organizers' hidden sequences; bacteria set from prior independent work) and do not reduce by construction to fitted inputs or self-definitions. The non-homophily measurement (Hadj≈0 on line graphs of candidate graphs) is a diagnostic of the input topology that motivates the edge-centric design; it is not used as a circular proof of tracking accuracy. Architecture equations (edge tokens Eq. 1, line-to-line bias Eq. 3, parental softmax Eq. 4, variable appearance Eq. 6, ILP) define a model that is then trained and scored; they do not equate the reported metrics to the inputs. Hyperparameters and ILP weights are tuned on held-out CTC splits. Self-citations (e.g. Ultrack) are ordinary prior work and not load-bearing uniqueness theorems. No fitted-parameter-as-prediction, ansatz-via-self-citation, or renaming-of-known-result steps appear. The paper is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 4 invented entities

This is an empirical systems paper. Load-bearing content is architectural inductive bias plus standard biological tracking constraints and public benchmarks, not a short list of physical postulates. Free parameters are the usual ML/ILP hyperparameters that affect reported numbers; axioms are domain constraints and the candidate-graph construction; invented entities are the method components themselves, each with ablation or external-benchmark handles.

free parameters (6)
  • Hidden dim C, heads H, node/edge depth Ln/Le
    Architecture size (C=288, H=4, Ln=Le=4) chosen by authors; capacity and compute of the reported model depend on them.
  • Candidate spatial threshold τ / k-NN
    τ=300 px (and k-NN sparsity) defines which true links can ever be recovered; central to the candidate-graph assumption.
  • ILP costs (λ, appearance/disappearance/division/node weights)
    λ=0.5, appearance 0.5, disappearance 0.25, division 0.25, node −10; tuned on half of CTC and used at inference for all reported scores.
  • Focal loss γ and division edge weight
    γ=3.5 and 3.5× weight on division edges; directly shapes training toward rare divisions.
  • Line-to-line bias init α_h and attractive/repulsive signs σ_h
    α_h init −5 (softplus near zero) and alternating σ_h ∈ {+1,−1} control the geometric attention prior that ablations credit for gains over plain edge Transformers.
  • Optimizer and schedule (Muon/Adam lrs, EMA, 50k steps)
    Hybrid Muon–Adam rates and cosine schedule are hand-chosen training knobs that affect final CTC/bacteria numbers.
axioms (5)
  • domain assumption Each cell has at most one parent and at most two daughters; solution is a binary forest in the candidate graph.
    Stated in Sec. 3 and encoded in parental softmax and ILP constraints (Eqs. 12–14).
  • domain assumption Spatially local candidate edges (distance ≤ τ, optional multi-frame Δt) contain the true associations given the input segmentations.
    Opening of Sec. 3; method never proposes links outside GC.
  • domain assumption Candidate tracking graphs are non-homophilic on the line graph (near-random label agreement among co-incident edges).
    Measured in Sec. 4.1 / Appendix C (Hadj≈0.01) and used to justify abandoning topology-driven aggregation.
  • standard math Standard Transformer attention, 3D RoPE with learnable frequencies/reflections, focal loss, and ILP solvers behave as in the cited literature.
    Background machinery from Vaswani et al., Su et al., Heo et al., Lin et al., Zhang et al. network-flow ILP lineage.
  • ad hoc to paper Hand-crafted 19-D geometric/intensity features per detection are sufficient inputs for competitive association without deep image encoders.
    Appendix A feature list; central to the ‘no pretrained encoder’ claim on CTC and bacteria.
invented entities (4)
  • HOCT edge-centric two-stage Transformer independent evidence
    purpose: Classify candidate links by edge–edge attention instead of node similarity or GNN message passing on GC.
    Core architecture (Sec. 3.1); independent handle via CTC/bacteria scores and edge-stage ablations.
  • Line-to-line distance attention bias (attractive/repulsive heads) independent evidence
    purpose: Inject geometric compatibility between candidate segments beyond midpoint RoPE.
    Sec. 3.2 Eq. 3; ablation Table 2 shows drop when removed (Edge-Transformer).
  • Multi-frame parental softmax with implicit no-parent and variable appearance cost independent evidence
    purpose: Normalize parents per target and Δt; derive appearance probabilities from the same logits.
    Sec. 3.3 Eqs. 4–6; Table 1 shows appearance probability helps CLB.
  • Two-pass tracklet ILP solver independent evidence
    purpose: Protect short-range associations before allowing long-range bridges.
    Sec. 3.3; Table 1 shows long links alone hurt unless two-pass is on.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Higher-Order Cell Tracking Transformer." pith.science (2026). https://pith.science/paper/NSPXAA4D

@misc{pith2026260711754,
  author       = {Pith},
  title        = {Pith review of: Higher-Order Cell Tracking Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NSPXAA4D}},
  note         = {Machine review of arXiv:2607.11754}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reconstructing lineages from live-imaging microscopy requires linking cell detections across time, including through cell divisions. A common approach is to construct a candidate graph and associate cell segmentations (nodes) across frames. However, these and other existing methods overlook two structural obstacles in candidate tracking graphs: (i) cell divisions entangle distinct lineage paths in the node embedding space, and (ii) edges sharing a node have near-random label agreement, so the candidate-graph topology carries no useful information for graph neural networks to aggregate. We propose the \textbf{Higher-Order Cell Tracking Transformer} (HOCT), an edge-centric architecture in which candidate cell links attend to one another under a 3D geometric prior, resolving both issues. Evaluated on the Cell Tracking Challenge and a bacteria division benchmark, HOCT achieves state-of-the-art results without deep pre-trained image encoders. Moreover, the proposed approach is easier to fine-tune, quickly reducing tracking errors by 59% with 400 annotations in a human-in-the-loop setting, outperforming LoRA fine-tuning of competing transformer baselines (6.75% improvement).

Figures

Figures reproduced from arXiv: 2607.11754 by Ilan Theodoro, Jord\~ao Bragantini, Lo\"ic A. Royer.

Figure 1
Figure 1. Figure 1: Motivation and architecture. (a) A candidate tracking graph and its ground-truth solution, where each color denotes a distinct cell trajectory (simple path in the lineage tree). The graph is non-homophilic: edges sharing a node have near-random label agreement, so graph topology carries no useful label information. (b) In node embedding approaches, divisions create connected manifolds that merge distinct l… view at source ↗
Figure 2
Figure 2. Figure 2: Relative AOGM reduction during incremental correction on bacteria validation sequences. Both panels show the percentage AOGM reduction from each system’s round-0 baseline, enabling direct comparison of correction efficiency despite different starting points (HOCT: ∼1400; Trackastra: ∼4950). (a) HOCT strategies reduce AOGM by up to 59% within 400 annotations, with uncertainty sampling achieving the steepest… view at source ↗
Figure 3
Figure 3. Figure 3: Per-sequence AOGM progression during incremental correction. Each row corresponds to one bacteria validation sequence. Left column: HOCT strategies; right column: Trackastra + At￾trackt strategies. The improvement percentages are consistent across sequences: HOCT reduces AOGM by 48–78%, while Trackastra + Attrackt achieves between ∼4% (150428-08) and ∼15% (151031-03), with LoRA + unsup the strongest varian… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 5 linked inside Pith

  1. [1]

    Leveraging learned object features for transformer-based cell tracking in mi- croscopy

    Cyril Achard. Leveraging learned object features for transformer-based cell tracking in mi- croscopy. Master’s thesis, École Polytechnique Fédérale de Lausanne, 2025

    2025

  2. [2]

    Graph neural network for cell tracking in microscopy videos

    Tal Ben-Haim and Tammy Riklin Raviv. Graph neural network for cell tracking in microscopy videos. InEuropean Conference on Computer Vision, pages 610–626. Springer, 2022

    2022

  3. [3]

    Tracking without bells and whistles

    Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. InIEEE International Conference on Computer Vision, pages 941–951, 2019

    2019

  4. [4]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. InIEEE International Conference on Image Processing, pages 3464–3468. IEEE, 2016

    2016

  5. [5]

    Beyond low-frequency information in graph convolutional networks

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency information in graph convolutional networks. InAAAI Conference on Artificial Intelligence, volume 35, pages 3950–3957, 2021

    2021

  6. [6]

    Swept confocally- aligned planar excitation (SCAPE) microscopy for high-speed volumetric imaging of behaving organisms.Nature Photonics, 9(2):113–119, 2015

    Matthew B Bouchard, Venkatakaushik V oleti, César S Mendes, Clay Lacefield, Wesley B Grueber, Richard S Mann, Randy M Bruno, and Elizabeth MC Hillman. Swept confocally- aligned planar excitation (SCAPE) microscopy for high-speed volumetric imaging of behaving organisms.Nature Photonics, 9(2):113–119, 2015

    2015

  7. [7]

    Large-scale multi-hypotheses cell tracking using ultrametric contours maps

    Jordão Bragantini, Merlin Lange, and Loïc Royer. Large-scale multi-hypotheses cell tracking using ultrametric contours maps. InEuropean Conference on Computer Vision, 2024

    2024

  8. [8]

    Ultrack: pushing the limits of cell tracking across biological scales.Nature Methods, pages 1–14, 2025

    Jordão Bragantini, Ilan Theodoro, Xiang Zhao, Teun APM Huijben, Eduardo Hirata-Miyasaki, Shruthi VijayKumar, Akilandeswari Balasubramanian, Tiger Lao, Richa Agrawal, Sheng Xiao, et al. Ultrack: pushing the limits of cell tracking across biological scales.Nature Methods, pages 1–14, 2025

    2025

  9. [9]

    Learning a neural solver for multiple object tracking

    Guillem Braso and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6246–6256. IEEE, 2020. doi: 10.1109/cvpr42600.2020.00628

    doi:10.1109/cvpr42600.2020.00628 2020

  10. [10]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean Conference on Computer Vision, pages 213–229. Springer, 2020

    2020

  11. [11]

    Long-term live imaging, cell identifica- tion and cell tracking in regenerating crustacean legs.eLife, 14:RP107534, 2025

    Ça˘grı Çevrim, Béryl Laplace-Builhé, Ko Sugawara, Maria Lorenza Rusciano, Nicolas Labert, Jacques Brocard, Alba Almazán, and Michalis Averof. Long-term live imaging, cell identifica- tion and cell tracking in regenerating crustacean legs.eLife, 14:RP107534, 2025

    2025

  12. [12]

    Lattice light-sheet microscopy: imaging molecules to embryos at high spatiotemporal resolution.Science, 346 (6208):1257998, 2014

    Bi-Chang Chen, Wesley R Legant, Kai Wang, Lin Shao, Daniel E Milkie, Michael W Davidson, Chris Janetopoulos, Xufeng S Wu, John A Hammer III, Zhe Liu, et al. Lattice light-sheet microscopy: imaging molecules to embryos at high spatiotemporal resolution.Science, 346 (6208):1257998, 2014

    2014

  13. [13]

    Mot20: A benchmark for multi object tracking in crowded scenes.arXiv preprint arXiv:2003.09003, 2020

    Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixé. Mot20: A benchmark for multi object tracking in crowded scenes.arXiv preprint arXiv:2003.09003, 2020

    Pith/arXiv arXiv 2003

  14. [14]

    TrackTour, 2025

    Draga Doncila Pop. TrackTour, 2025. URL https://github.com/DragaDoncila/ tracktour

    2025

  15. [15]

    Optical pooled screens in human cells

    David Feldman, Avtar Singh, Jonathan L Schmid-Burgk, Rebecca J Carlson, Anja Mezger, Anthony J Garrity, Feng Zhang, and Paul C Blainey. Optical pooled screens in human cells. Cell, 179(3):787–799, 2019

    2019

  16. [16]

    Trackastra: Transformer-based cell tracking for live-cell microscopy

    Benjamin Gallusser and Martin Weigert. Trackastra: Transformer-based cell tracking for live-cell microscopy. InEuropean Conference on Computer Vision, 2024

    2024

  17. [17]

    Trackastra: code and pretrained weights, 2024

    Benjamin Gallusser and Martin Weigert. Trackastra: code and pretrained weights, 2024. URL https://github.com/weigertlab/trackastra. 10

    2024

  18. [18]

    ASCENT: Annotation-free self-supervised contrastive embeddings for 3d neuron tracking in fluorescence microscopy

    Haejun Han and Hang Lu. ASCENT: Annotation-free self-supervised contrastive embeddings for 3d neuron tracking in fluorescence microscopy. InIEEE International Conference on Computer Vision, pages 14676–14687, 2025

    2025

  19. [19]

    Mpm: Joint representation of motion and position map for cell tracking

    Junya Hayashida, Kazuya Nishimura, and Ryoma Bise. Mpm: Joint representation of motion and position map for cell tracking. InIEEE Conference on Computer Vision and Pattern Recognition, pages 3823–3832, 2020

    2020

  20. [20]

    Rotary position embedding for vision transformer

    Byeongho Heo, Song Park, Dongyoon Han, and Sangdoo Yun. Rotary position embedding for vision transformer. InEuropean Conference on Computer Vision, pages 289–305. Springer, 2024

    2024

  21. [21]

    Micrographia

    Robert Hooke. Micrographia. InLiterature and Science, 1660-1834, Part II vol 5, pages 1–14. Routledge, 1665

  22. [22]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022

    2022

  23. [23]

    Robust single-particle tracking in live-cell time-lapse sequences

    Khuloud Jaqaman, Dinah Loerke, Marcel Mettlen, Hirotaka Kuwata, Sergio Grinstein, Sandra L Schmid, and Gaudenz Danuser. Robust single-particle tracking in live-cell time-lapse sequences. Nature Methods, 5(8):695–702, 2008

    2008

  24. [24]

    Muon: An optimizer for hidden layers in neural networks, 2024

    Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. Muon: An optimizer for hidden layers in neural networks, 2024. URL https://kellerjordan.github.io/posts/muon/

    2024

  25. [25]

    Cell tracking according to biolog- ical needs-strong mitosis-aware multi-hypothesis tracker with aleatoric uncertainty.IEEE Transactions on Medical Imaging, 2025

    Timo Kaiser, Maximilian Schier, and Bodo Rosenhahn. Cell tracking according to biolog- ical needs-strong mitosis-aware multi-hypothesis tracker with aleatoric uncertainty.IEEE Transactions on Medical Imaging, 2025

    2025

  26. [26]

    Reconstruction of zebrafish early embryonic development by scanned light sheet microscopy.Science, 322 (5904):1065–1069, 2008

    Philipp J Keller, Annette D Schmidt, Joachim Wittbrodt, and Ernst HK Stelzer. Reconstruction of zebrafish early embryonic development by scanned light sheet microscopy.Science, 322 (5904):1065–1069, 2008

    2008

  27. [27]

    The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    Harold W Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    1955

  28. [28]

    Attrackt, 2025

    Manan Lalit and Jan Funke. Attrackt, 2025. URL https://github.com/funkelab/ attrackt

    2025

  29. [29]

    Attrackt experiments, 2025

    Manan Lalit and Jan Funke. Attrackt experiments, 2025. URL https://github.com/ funkelab/attrackt_experiments

    2025

  30. [30]

    An investigation of unsupervised cell tracking and interactive fine- tuning

    Manan Lalit and Jan Funke. An investigation of unsupervised cell tracking and interactive fine- tuning. InIEEE International Conference on Computer Vision Workshops, pages 5792–5800, 2025

    2025

  31. [31]

    A multimodal zebrafish developmental atlas reveals the state-transition dynamics of late-vertebrate pluripotent axial progenitors.Cell, 187(23):6742–6759, 2024

    Merlin Lange, Alejandro Granados, Shruthi VijayKumar, Jordão Bragantini, Sarah Ancheta, Yang-Joon Kim, Sreejith Santhosh, Michael Borja, Hirofumi Kobayashi, Erin McGeever, et al. A multimodal zebrafish developmental atlas reveals the state-transition dynamics of late-vertebrate pluripotent axial progenitors.Cell, 187(23):6742–6759, 2024

    2024

  32. [32]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InIEEE International Conference on Computer Vision, pages 2980–2988, 2017

    2017

  33. [33]

    Rethinking rope: A mathematical blueprint for n-dimensional positional embedding

    Haiping Liu, Lijing Lin, Jingyuan Sun, Zhegong Shangguan, Mauricio A Alvarez, and Hong- peng Zhou. Rethinking rope: A mathematical blueprint for n-dimensional positional embedding. arXiv preprint arXiv:2504.06308, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    PhD thesis, KTH Royal Institute of Technology, 2016

    Klas EG Magnusson.Segmentation and tracking of cells and particles in time-lapse microscopy. PhD thesis, KTH Royal Institute of Technology, 2016. 11

    2016

  35. [35]

    Automated reconstruction of whole-embryo cell lineages by learning from sparse annotations.Nature Biotechnology, pages 1–6, 2022

    Caroline Malin-Mayor, Peter Hirsch, Leo Guignard, Katie McDole, Yinan Wan, William C Lemon, Dagmar Kainmueller, Philipp J Keller, Stephan Preibisch, and Jan Funke. Automated reconstruction of whole-embryo cell lineages by learning from sparse annotations.Nature Biotechnology, pages 1–6, 2022

    2022

  36. [36]

    The cell tracking challenge: 10 years of objective benchmarking.Nature Methods, pages 1–11, 2023

    Martin Maška, Vladimír Ulman, Pablo Delgado-Rodriguez, Estibaliz Gómez-de Mariscal, Tereza Neˇcasová, Fidel A Guerrero Peña, Tsang Ing Ren, Elliot M Meyerowitz, Tim Scherr, Katharina Löffler, et al. The cell tracking challenge: 10 years of objective benchmarking.Nature Methods, pages 1–11, 2023

    2023

  37. [37]

    Cell tracking accuracy measurement based on comparison of acyclic oriented graphs.PloS one, 10(12):e0144959, 2015

    Pavel Matula, Martin Maška, Dmitry V Sorokin, Petr Matula, Carlos Ortiz-de Solórzano, and Michal Kozubek. Cell tracking accuracy measurement based on comparison of acyclic oriented graphs.PloS one, 10(12):e0144959, 2015

    2015

  38. [38]

    Track- former: Multi-object tracking with transformers

    Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Track- former: Multi-object tracking with transformers. InIEEE Conference on Computer Vision and Pattern Recognition, pages 8844–8854, 2022

    2022

  39. [39]

    Quasi- dense similarity learning for multiple object tracking

    Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi- dense similarity learning for multiple object tracking. InIEEE Conference on Computer Vision and Pattern Recognition, pages 164–173, 2021

    2021

  40. [40]

    Geometric deep learning reveals the spatiotemporal features of microscopic motion

    Jesús Pineda, Benjamin Midtvedt, Harshith Bachimanchi, Sergio Noé, Daniel Midtvedt, Gio- vanni V olpe, and Carlo Manzo. Geometric deep learning reveals the spatiotemporal features of microscopic motion. 5(1):71–82, 2023

    2023

  41. [41]

    A critical look at the evaluation of GNNs under heterophily: Are we re- ally making progress? InInternational Conference on Learning Representations, 2023

    Oleg Platonov, Denis Kuznedelev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. A critical look at the evaluation of GNNs under heterophily: Are we re- ally making progress? InInternational Conference on Learning Representations, 2023

    2023

  42. [42]

    Needles in the haystack-rapid error detection for cell tracking solutions

    Draga Doncila Pop, Pierre Le Bodic, and Juan Nunez-Iglesias. Needles in the haystack-rapid error detection for cell tracking solutions. InInt. Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2025

    2025

  43. [43]

    Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free.arXiv preprint arXiv:2505.06708, 2025

    Zihan Qiu, Zekun Wang, Bo Zheng, Zeyu Huang, Kaiyue Wen, Songlin Yang, Rui Men, Le Yu, Fei Huang, Suozhi Huang, et al. Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free.arXiv preprint arXiv:2505.06708, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Particle tracking in biological images with optical-flow enhanced kalman filtering

    Raphael Reme, Alasdair Newson, Elsa Angelini, Jean-Christophe Olivo-Marin, and Thibault Lagache. Particle tracking in biological images with optical-flow enhanced kalman filtering. In Int. Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2024

    2024

  45. [45]

    Pathfinder discovery networks for neural message passing

    Benedek Rozemberczki, Peter Englert, Amol Kapoor, Martin Blais, and Bryan Perozzi. Pathfinder discovery networks for neural message passing. InProceedings of the Web Confer- ence 2021, pages 2547–2558, 2021

    2021

  46. [46]

    Conservation tracking

    Martin Schiegg, Philipp Hanslovsky, Bernhard X Kausler, Lars Hufnagel, and Fred A Hamprecht. Conservation tracking. InIEEE Conference on Computer Vision and Pattern Recognition, pages 2928–2935, 2013

    2013

  47. [47]

    Smart lattice light-sheet microscopy for imaging rare and complex cellular events.Nature Methods, pages 1–10, 2024

    Yu Shi, Jimmy S Tabet, Daniel E Milkie, Timothy A Daugird, Chelsea Q Yang, Alex T Ritter, Andrea Giovannucci, and Wesley R Legant. Smart lattice light-sheet microscopy for imaging rare and complex cellular events.Nature Methods, pages 1–10, 2024

    2024

  48. [48]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  49. [49]

    Tracking cell lineages in 3d by incremental deep learning.eLife, 11:e69380, 2022

    Ko Sugawara, Ça˘grı Çevrim, and Michalis Averof. Tracking cell lineages in 3d by incremental deep learning.eLife, 11:e69380, 2022

    2022

  50. [50]

    Network flow integer programming to track elliptical cells in time-lapse sequences.IEEE Transactions on Medical Imaging, 36(4):942–951, 2016

    Engin Türetken, Xinchao Wang, Carlos J Becker, Carsten Haubold, and Pascal Fua. Network flow integer programming to track elliptical cells in time-lapse sequences.IEEE Transactions on Medical Imaging, 36(4):942–951, 2016. 12

    2016

  51. [51]

    An objective comparison of cell-tracking algorithms.Nature Methods, 14(12):1141–1152, 2017

    Vladimír Ulman, Martin Maška, Klas EG Magnusson, Olaf Ronneberger, Carsten Haubold, Nathalie Harder, Pavel Matula, Petr Matula, David Svoboda, Miroslav Radojevic, et al. An objective comparison of cell-tracking algorithms.Nature Methods, 14(12):1141–1152, 2017

    2017

  52. [52]

    Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

    2017

  53. [53]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. InInternational Conference on Learning Representations, 2018

    2018

  54. [54]

    Sam-med3d: a vision foundation model for general-purpose segmentation on volumetric medical images.IEEE Transactions on Neural Networks and Learning Systems, 2025

    Haoyu Wang, Sizheng Guo, Jin Ye, Zhongying Deng, Junlong Cheng, Tianbin Li, Jianpin Chen, Yanzhou Su, Ziyan Huang, Yiqing Shen, et al. Sam-med3d: a vision foundation model for general-purpose segmentation on volumetric medical images.IEEE Transactions on Neural Networks and Learning Systems, 2025

    2025

  55. [55]

    Simple online and realtime tracking with a deep association metric

    Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. InIEEE International Conference on Image Processing, pages 3645–3649. IEEE, 2017

    2017

  56. [56]

    Daxi—high-resolution, large imaging volume and multi-view single-objective light-sheet mi- croscopy.Nature Methods, 19(4):461–469, 2022

    Bin Yang, Merlin Lange, Alfred Millett-Sikking, Xiang Zhao, Jordão Bragantini, Shruthi Vi- jay Kumar, Mason Kamb, Rafael Gómez-Sjöberg, Ahmet Can Solak, Wanpeng Wang, et al. Daxi—high-resolution, large imaging volume and multi-view single-objective light-sheet mi- croscopy.Nature Methods, 19(4):461–469, 2022

    2022

  57. [57]

    Global data association for multi-object tracking using network flows

    Li Zhang, Yuan Li, and Ramakant Nevatia. Global data association for multi-object tracking using network flows. InIEEE Conference on Computer Vision and Pattern Recognition, pages 1–8. IEEE, 2008

    2008

  58. [58]

    ByteTrack: Multi-object tracking by associating every detection box

    Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. ByteTrack: Multi-object tracking by associating every detection box. InEuropean Conference on Computer Vision, pages 1–21. Springer, 2022

    2022

  59. [59]

    Graph neural networks for graphs with heterophily: A survey.arXiv preprint arXiv:2202.07082, 2022

    Xin Zheng, Yi Wang, Yixin Liu, Ming Li, Miao Zhang, Di Jin, Philip S Yu, and Shirui Pan. Graph neural networks for graphs with heterophily: A survey.arXiv preprint arXiv:2202.07082, 2022

    Pith/arXiv arXiv 2022

  60. [60]

    CELLECT: contrastive embedding learning for large-scale efficient cell tracking.Nature Methods, pages 1–12, 2025

    Hongyu Zhou, Seonghoon Kim, Zhifeng Zhao, Jiaqi Fan, Wen Huang, Xinghua Sui, Lizhi Shao, Haoran An, Jing-Ren Zhang, Jiamin Wu, et al. CELLECT: contrastive embedding learning for large-scale efficient cell tracking.Nature Methods, pages 1–12, 2025

    2025

  61. [61]

    Medical sam 2: Segment medical images as video via segment anything model 2.arXiv preprint arXiv:2408.00874, 2024

    Jiayuan Zhu, Abdullah Hamdi, Yunli Qi, Yueming Jin, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2.arXiv preprint arXiv:2408.00874, 2024

    Pith/arXiv arXiv 2024

  62. [62]

    MLP_E”/“MLP_D

    Zi Lab. zTrack4CTC, 2024. URLhttps://github.com/Zi-Lab/zTrack4CTC. A Implementation details Architecture.The full architecture specifications are: input dimension d=19, hidden dimension C=288, H=4 attention heads (head dimension D=72), MLP expansion factor 2 (hidden MLP dimension 576), attention dropout 0.05, MLP dropout 0.0, gated attention [43] in all l...

    2024

  63. [63]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

This paper was first reviewed by grok-4.5 on July 14, 2026.