Pith. sign in

REVIEW 40 references

Single-shot Star-convex Polygon-based Instance Segmentation for Spatially-correlated Biomedical Objects

T0 review · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A branched StarDist architecture with a within-boundary penalty segments nested biomedical objects such as nuclei in cells and plaques in wells in one shot, with a new joint true-positive metric.

desk verdict A useful single-shot extension of StarDist for nested objects, but the WBR penalty's gradient path and the JTPR metric need fixing before the claims hold. read the letter →

arxiv 2504.12078 v1 pith:PP5EXVYA submitted 2025-04-16 cs.CV q-bio.QM

Trina De , Adrian Urbanski , Artur Yakimovich This is my paper · ORCID
classification cs.CVq-bio.QM
keywords instanceobjectssegmentationbiomedicalhsd-wbrnestedachievearchitectures
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Many biomedical images contain objects that sit inside other objects: cell nuclei live inside the cell body, and virus plaques grow inside the wells of a culture plate. Most segmentation systems detect each object type separately, which means running the model twice. This paper instead builds a single network with two output heads on a shared backbone. Both heads are based on StarDist, a method that represents objects as star-shaped polygons. The shared backbone lets the model use the fact that one object contains the other as a hint while learning.

The new part is a penalty called Within Boundary Regularisation. It checks, pixel by pixel, whether the predicted inner objects leak outside the predicted outer objects. If they do, the loss becomes larger. In principle this should push the model to keep nuclei inside cells and plaques inside wells. The authors also propose a new evaluation score, JTPR, which counts objects as successful only when both the inner object and its containing outer object are found together.

The method is tested on two published datasets of fluorescence microscopy and plate photographs. On the standard IoUR and AP metrics the new models are competitive but not always best. On their own JTPR score, the models often lead, though on one dataset the plain StarDist baseline scores higher on the inner object. The paper currently provides no code and no error bars, and it does not explain how the penalty is differentiated during training, which is a key detail. The idea is useful and likely to be built upon, but the evidence as written leaves room for doubt.

Extended reading notes

Core claim

The central assertion is that the proposed branched StarDist architectures HSD and HSD-WBR achieve nested instance segmentation of spatially correlated biomedical objects in a single shot, and that HSD-WBR, through the Within Boundary Regularisation penalty, outperforms StarDist and Cellpose on the proposed Joint TP rate (JTPR) criterion while remaining competitive on IoUR and AP (Abstract; Section 1; Tables 1-2).

Load-bearing premise

The WBR penalty, as defined in Equation 5 on predicted instance masks, must provide a usable training gradient. If the masks come from the non-differentiable NMS and polygon construction pipeline, the penalty term Lambda is a constant with respect to network parameters and cannot change learning, making HSD-WBR equivalent to HSD during training. The paper never states how gradients are obtained for Lambda, so the central mechanism may be inert (Section 3.3, Equation 5).

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are the regularization weights and the IoU threshold, all tuned on the evaluation data. The key unstated assumption is the differentiability of the penalty.

free parameters (5)
  • lambda1 = 0.2
    Weight on StarDist distance loss; selected via ablation in Appendix 8.2.
  • lambda2 = 0.0001
    Weight penalizing radial distance predictions at background pixels; selected via ablation.
  • lambda3 = 1
    Weight on the WBR penalty; selected via ablation on validation/test data.
  • epsilon = 1e-7
    Small constant in the WBR denominator; set by hand to bound the penalty.
  • IoU threshold tau = optimal per model and metric on test set
    Headline IoUR, AP, and JTPR values are reported at the per-model optimal tau, which is a form of test-set fitting.
assumptions (3)
  • domain assumption Both object classes in each dataset are well approximated by star-convex polygons with K=32 radial directions.
    The method inherits StarDist's star-convex representation; K=32 is set by the authors with no quantitative justification.
  • domain assumption Nuclei are always inside cytoplasm, and plaques are always inside wells; the containment relation is strict and known ahead of time.
    The WBR penalty encodes this prior; the paper's claims depend on the relation holding in all test images.
  • ad hoc to paper The WBR penalty computed on predicted instance masks can be differentiated with respect to network parameters.
    No gradient path is described through the NMS and polygon construction step, so the penalty may not affect training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Single-shot Star-convex Polygon-based Instance Segmentation for Spatially-correlated Biomedical Objects." pith.science (2026). https://pith.science/paper/PP5EXVYA

@misc{pith2026250412078,
  author       = {Pith},
  title        = {Pith review of: Single-shot Star-convex Polygon-based Instance Segmentation for Spatially-correlated Biomedical Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PP5EXVYA}},
  note         = {Machine review of arXiv:2504.12078}
}
abstract

Biomedical images often contain objects known to be spatially correlated or nested due to their inherent properties, leading to semantic relations. Examples include cell nuclei being nested within eukaryotic cells and colonies growing exclusively within their culture dishes. While these semantic relations bear key importance, detection tasks are often formulated independently, requiring multi-shot analysis pipelines. Importantly, spatial correlation could constitute a fundamental prior facilitating learning of more meaningful representations for tasks like instance segmentation. This knowledge has, thus far, not been utilised by the biomedical computer vision community. We argue that the instance segmentation of two or more categories of objects can be achieved in parallel. We achieve this via two architectures HydraStarDist (HSD) and the novel (HSD-WBR) based on the widely-used StarDist (SD), to take advantage of the star-convexity of our target objects. HSD and HSD-WBR are constructed to be capable of incorporating their interactions as constraints into account. HSD implicitly incorporates spatial correlation priors based on object interaction through a joint encoder. HSD-WBR further enforces the prior in a regularisation layer with the penalty we proposed named Within Boundary Regularisation Penalty (WBR). Both architectures achieve nested instance segmentation in a single shot. We demonstrate their competitiveness based on $IoU_R$ and AP and superiority in a new, task-relevant criteria, Joint TP rate (JTPR) compared to their baseline SD and Cellpose. Our approach can be further modified to capture partial-inclusion/-exclusion in multi-object interactions in fluorescent or brightfield microscopy or digital imaging. Finally, our strategy suggests gains by making this learning single-shot and computationally efficient.

Figures

Figures reproduced from arXiv: 2504.12078 by the authors.

Figure 1
Figure 1. StarDist-based instance segmentation architectures (Section 3). Architecture diagrams of the instance [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Limitations of approach in [Delong and Boykov, 2009]. Here object 2 bounds objects 1 and 3 that may either [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. HeLaCytoNuc [De et al., 2024a] dataset (Subsection 4.1). (a) Fluorescence micrograph with cytoplasm and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: VACVPlaque [De et al., 2024b] dataset (Subsection 4.2). (a) RGB mobile photographs of plaques within [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: HeLaCytoNuc Single-shot instance segmentation results (Sections 4 and 6), Predicted and Ground Truth (GT) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: VACVPlaque Single-shot instance segmentation results (Sections 4 and 6), Predicted and Ground Truth (GT) [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: HeLaCytoNuc Two-shot instance segmentation results (Sections 4 and 6), Predicted and Ground Truth (GT) [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: VACVPlaque Two-shot instance segmentation results (Sections 2 and 4), Predicted and Ground Truth (GT) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 23 canonical work pages

  1. [1]

    Cell detection with star-convex polygons

    Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell detection with star-convex polygons. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2018 - 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II , pages 265--273, 2018. doi:10.1007/978-3-030-00934-2_30

    doi:10.1007/978-3-030-00934-2_30 2018

  2. [2]

    Cellpose: a generalist algorithm for cellular segmentation

    Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18 0 (1): 0 100--106, 2021. doi:10.1038/s41592-020-01018-x

    doi:10.1038/s41592-020-01018-x 2021

  3. [3]

    Carpenter, Thouis R

    Anne E. Carpenter, Thouis R. Jones, M.R. Lamprecht, Colin Clarke, In Han Kang, Ola Friman, David A. Guertin, Joo Han Chang, Robert A. Lindquist, Jason Moffat, et al. Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. Genome biology, 7: 0 1--11, 2006. doi:10.1186/gb-2006-7-10-r100

    doi:10.1186/gb-2006-7-10-r100 2006

  4. [4]

    Humphries, P

    M.P. Humphries, P. Maxwell, and M. Salto-Tellez. Qupath: The global impact of an open source digital pathology system. Computational and Structural Biotechnology Journal, 19: 0 852--859, 2021. doi:10.1038/s41598-017-17204-5

    doi:10.1038/s41598-017-17204-5 2021

  5. [5]

    Vaccinia virus vaccines: past, present and future

    Bertram L Jacobs, Jeffrey O Langland, Karen V Kibler, Karen L Denzler, Stacy D White, Susan A Holechek, Shukmei Wong, Trung Huynh, and Carole R Baskin. Vaccinia virus vaccines: past, present and future. Antiviral Res., 84 0 (1): 0 1--13, October 2009

    work page 2009

  6. [6]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234--241, 2015. doi:10.1007/978-3-319-24574-4_28

    doi:10.1007/978-3-319-24574-4_28 2015

  7. [7]

    Isoodl: Instance segmentation of overlapping biological objects using deep learning

    Anton Bohm, Annekathrin Ucker, Tim Jager, Olaf Ronneberger, and Thorsten Falk. Isoodl: Instance segmentation of overlapping biological objects using deep learning. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), pages 1225--1229. IEEE, 2018. doi:10.1109/ISBI.2018.8363792

    arXiv 2018

  8. [8]

    Isoov2dl - semantic instance segmentation of touching and overlapping objects

    Anton Bohm, Annekathrin Ucker, Tim Jager, Olaf Ronneberger, and Thorsten Falk. Isoov2dl - semantic instance segmentation of touching and overlapping objects. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pages 343--347. IEEE, 2019. doi:10.1109/ISBI.2019.8759334

    arXiv 2019

Show all 40 references
  1. [9]

    A novel medical image segmentation approach by using multi-branch segmentation network based on local and global information synchronous learning

    Shangzhu Jin, Sheng Yu, Jun Peng, Hongyi Wang, and Yan Zhao. A novel medical image segmentation approach by using multi-branch segmentation network based on local and global information synchronous learning. Nature Scientific Reports, 13 0 (1): 0 6762, 2023. doi:10.1038/s41598...

    2023 doi

  2. [10]

    Double branch parallel network for segmentation of buildings and waters in remote sensing images

    Jing Chen, Min Xia, Dehao Wang, and Haifeng Lin. Double branch parallel network for segmentation of buildings and waters in remote sensing images. Remote Sensing, 15 0 (6), 2023 a . doi:10.3390/rs15061536

    2023 doi

  3. [11]

    Globally optimal segmentation of multi-region objects

    A Delong and Y Boykov. Globally optimal segmentation of multi-region objects. In textit 2009 IEEE 12th International Conference on Computer Vision , pages 285--292. 2009

    2009

  4. [12]

    Fast approximate energy minimization with label costs

    A Delong, A Osokin, H N Isack, and Y Boykov. Fast approximate energy minimization with label costs. In textit 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition , pages 2173--2180. 2010

    2010

  5. [13]

    Computer vision approaches to medical image analysis

    M Sonka, editor. Computer vision approaches to medical image analysis. textit\ Springer\ , Berlin, Germany, 2006

    2006

  6. [14]

    Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for Multi-Human parsing

    J Zhao, J Li, Y Cheng, T Sim, S Yan, and J Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for Multi-Human parsing. In textit Proceedings of the 26th ACM international conference on Multimedia , pages 792--800. 2018

    2018

  7. [15]

    The distribution of the flora in the alpine zone.1

    Paul Jaccard. The distribution of the flora in the alpine zone.1. New Phytologist, 11 0 (2): 0 37--50, 1912. doi:10.1111/j.1469-8137.1912.tb05611.x

    1912

  8. [16]

    A delay metric for video object detection: What average precision fails to tell

    Huizi Mao, Xiaodong Yang, and Bill Dally. A delay metric for video object detection: What average precision fails to tell. In 2019 IEEE/CVF International Conference on Computer Vision ( ICCV ) . IEEE, October 2019

    2019

  9. [17]

    Robust and decomposable average precision for image retrieval

    E Ramzi, N Thome, C Rambour, N Audebert, and X Bitot. Robust and decomposable average precision for image retrieval. In Neural Information Processing Systems. 2021

    2021

  10. [18]

    SortedAP : Rethinking evaluation metrics for instance segmentation

    Long Chen, Yuli Wu, Johannes Stegmaier, and Dorit Merhof. SortedAP : Rethinking evaluation metrics for instance segmentation. In 2023 IEEE/CVF International Conference on Computer Vision Workshops ( ICCVW ) , pages 3925--3931. IEEE, October 2023 b

    2023

  11. [19]

    Allen Kent, M. M. Berry, Fred U. Luehrs, and James W. Perry. Machine literature searching viii. operational criteria for designing information retrieval systems. American Documentation, 6: 0 93--101, 1955. doi:10.1002/asi.5090060209

    1955 doi

  12. [20]

    Gismondi

    Marco Cacciabue, Anabella Currá, and Maria I. Gismondi. Viralplaque: a fiji macro for automated assessment of viral plaque statistics. PeerJ, 7: 0 e7729, 2019. doi:10.7717/peerj.7729

    2019 doi

  13. [21]

    Machine-learning-based automated quantification machine for virus plaque assay counting

    Gridsada Phanomchoeng, Chayatorn Kukiattikoon, Suphanut Plengkham, Siwaporn Boonyasuppayakorn, Saran Salakij, Suvit Poomrittigul, and Lunchakorn Wuttisittikulkij. Machine-learning-based automated quantification machine for virus plaque assay counting. PeerJ Computer Science, 8...

    2022 doi

  14. [22]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521 0 (7553): 0 436--444, 2015. doi:10.1038/nature14539

    2015 doi

  15. [23]

    James R. Munkres. Topology. Featured Titles for Topology. Prentice Hall, Incorporated, 2000. ISBN 978-0-13-181629-9

    2000

  16. [24]

    Optimizing star-convex functions

    Jasper Lee and Paul Valiant. Optimizing star-convex functions. In 2016 IEEE 57th Annual Symposium on Foundations of Computer Science, pages 603--614, 2016. doi:10.1109/FOCS.2016.71

    2016 doi

  17. [25]

    A computational approach to edge detection

    John Canny. A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell., PAMI-8 0 (6): 0 679--698, November 1986

    1986

  18. [26]

    A mathematical theory of communication

    Claude Elwood Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27: 0 379--423, 1948. doi:http://dx.doi.org/10.1002/j.1538-7305.1948.tb01338.x

    1948

  19. [27]

    Anomaly Detection and Complex Event Processing over IoT Data Streams With Application to eHealth and Patient Data Monitoring; Mean Absolute Error

    Patrick Schneider and Fatos Xhafa. Anomaly Detection and Complex Event Processing over IoT Data Streams With Application to eHealth and Patient Data Monitoring; Mean Absolute Error. Academic Press, 2022. ISBN 978-0-12-823818-9. doi:10.1016/C2020-0-00589-X

    2022 doi

  20. [28]

    Splinedist: Automated cell segmentation with spline curves

    Soham Mandal and Virginie Uhlmann. Splinedist: Automated cell segmentation with spline curves. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1082--1086, 2021. doi:10.1109/ISBI48211.2021.9433928

    2021

  21. [29]

    Set Theory: An Introduction to Independence Proofs

    Kenneth Kunen. Set Theory: An Introduction to Independence Proofs. North-Holland, 1980. ISBN 0444868399, 0444564020. doi:10.2307/2274070

    1980 doi

  22. [30]

    HeLaCytoNuc: fluorescence microscopy dataset with segmentation masks for cell nuclei and cytoplasm , June 2024 a

    Trina De, Adrian Urbanski, Subasini Thangamani, Maria Wyrzykowska, and Artur Yakimovich. HeLaCytoNuc: fluorescence microscopy dataset with segmentation masks for cell nuclei and cytoplasm , June 2024 a . URL https://doi.org/10.14278/rodare.3001

    2024 doi

  23. [31]

    Lee R. Dice. Measures of the amount of ecologic association between species. Ecology, Ecological Society of America, 26 0 (3): 0 297--302, 1945. doi:10.2307/1932409

    1945 doi

  24. [32]

    T. A. Sorensen. A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on Danish commons, volume 5(4). Kongelige Danske Videnskabernes Selskab, 1948

    1948

  25. [33]

    VACVPlaque: mobile photography of Vaccinia virus plaque assay with segmentation masks , June 2024 b

    Trina De, Adrian Urbanski, Subasini Thangamani, Maria Wyrzykowska, and Artur Yakimovich. VACVPlaque: mobile photography of Vaccinia virus plaque assay with segmentation masks , June 2024 b . URL https://doi.org/10.14278/rodare.3003

    2024 doi

  26. [34]

    TensorFlow : Large-scale machine learning on heterogeneous systems, 2015

    Mart\' i n Abadi et al. TensorFlow : Large-scale machine learning on heterogeneous systems, 2015. URL https://www.tensorflow.org/. Software available from tensorflow.org

    2015

  27. [35]

    PyTorch : An imperative style, High-Performance deep learning library

    A Paszke. PyTorch : An imperative style, High-Performance deep learning library. In textit Proceedings of the 33rd International Conference on Neural Information Processing Systems , volume 721, pages 8026--8037. Curran Associates Inc, Red Hook, NY, USA, Article

  28. [36]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. URL http://arxiv.org/abs/1412.6980

    2015 arXiv

  29. [37]

    Pytorch 1.9.0 documentation: Reducelronplateau, 2016

    Adam Paszke, Sam Gross, Soumith Chintala, and Gregory Chanan. Pytorch 1.9.0 documentation: Reducelronplateau, 2016. URL https://pytorch.org/docs/stable/generated/torch.optim.lr\_scheduler.ReduceLROnPlateau.html

    2016

  30. [38]

    Bertsekas

    Dimitri P. Bertsekas. Incremental gradient, subgradient, and proximal methods for convex optimization: A survey. In Optimization for Machine Learning. The MIT Press, 2011

    2011

  31. [39]

    Lightweight deep learning for resource-constrained environments: A survey

    Hou-I Liu, Marco Galindo, Hongxia Xie, Lai-Kuan Wong, Hong-Han Shuai, Yung-Hui Li, and Wen-Huang Cheng. Lightweight deep learning for resource-constrained environments: A survey. ACM Comput. Surv., 56 0 (10): 0 1--42, October 2024

    2024

  32. [40]

    The computational limits of deep learning

    Neil C Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F Manso. The computational limits of deep learning. 2020. doi:10.48550/arXiv.2007.05558

    2020 doi arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.