CoChat
Sign Up
Cited by 1,381Cell Image Analysis TechniquesAdvanced Neural Network ApplicationsMedical Image Segmentation Techniques

Cellpose 2.0: how to train your own model

Open access
Pachitariu, Marius, Stringer, Carsen·2022·Nature Methods
DOI 10.1038/s41592-022-01663-4PMC PMC9718665

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Pretrained neural network models for biological segmentation can provide good out-of-the-box results for many image types. However, such models do not allow users to adapt the segmentation style to their specific needs and can perform suboptimally for test images that are very different from the training images. Here we introduce Cellpose 2.0, a new package that includes an ensemble of diverse pretrained models as well as a human-in-the-loop pipeline for rapid prototyping of new custom models. We show that models pretrained on the Cellpose dataset can be fine-tuned with only 500–1,000 user-annotated regions of interest (ROI) to perform nearly as well as models trained on entire datasets with up to 200,000 ROI. A human-in-the-loop approach further reduced the required user annotation to 100–200 ROI, while maintaining high-quality segmentations. We provide software tools such as an annotation graphical user interface, a model zoo and a human-in-the-loop pipeline to facilitate the adoption of Cellpose 2.0.

AI snapshot

Written from the paper’s full text.

TL;DR

Cellpose 2.0 introduces an ensemble of pretrained segmentation models plus a human-in-the-loop GUI for rapidly fine-tuning custom models with minimal annotation. It shows that 500–1,000 user-annotated ROI offline, or 100–200 ROI with human-in-the-loop, can approach models trained on up to 200,000 ROI.

Key findings

  • An ensemble of nine style-clustered Cellpose models significantly outperformed a single global model, with average precision improvements of 0.01–0.06, largest at higher IoU thresholds and for the most diverse image class (CPx).
  • Fine-tuning pretrained Cellpose with 500–1,000 ROI per category achieved near-maximal accuracy; pretrained models reached average precision 0.73 at 426 training ROI versus 0.68 from scratch, and outperformed the Mesmer model starting at 1,000 training ROI (two 512 × 512 images) despite Mesmer being trained on up to 200,000 ROI.
  • Human-in-the-loop reduced annotation burden: the online/pretrained model required 167 total ROI versus 663 ROI for a standard offline approach, and 100–200 manually segmented ROI were sufficient for near-maximal accuracy in 1–2 hours per category.
  • In the breast vectra TissueNet example, average precision improved from 0.36 to 0.68 with four 256 × 256 images, then to 0.76 with 16 images and with all 524 images; in LiveCell A172, performance improved dramatically with 504 training ROI and reached maximum at 81,832 ROI.

Methods

  • Design: Cellpose 2.0 package with a model zoo, annotation GUI, and human-in-the-loop retraining pipeline; comparisons included pretrained versus scratch initialization and online versus offline annotation.
  • Data: Cellpose cyto dataset (540 images), TissueNet (2,601 training and 1,249 test images; 13 categories used, excluding lung mibi), and LiveCell (3,188 training and 1,516 test images; eight cell lines); overlapping LiveCell ROI were reassigned or removed.
  • Analysis: style vectors from Cellpose were clustered with the Leiden algorithm into nine classes, with separate models trained per cluster; segmentation quality was measured by average precision AP = TP/(TP+FP+FN) across IoU thresholds, with training via stochastic gradient descent (300 epochs offline, 100 epochs online).

Limitations

  • Stated: the LiveCell baseline model was trained to report overlapping ROI, so overlaps were removed before comparison; the within-human upper bound is by construction higher than any across-human upper bound; the authors note their conclusions may seem at odds with prior large-dataset studies due to scope, faster saturation of pretrained models, Cellpose performance, focus on tens-to-hundreds of ROI, augmentations, and a different average precision metric.
  • Evident: human-in-the-loop evaluation used the same human annotator and only two image categories (TissueNet breast vectra and LiveCell A172), so across-annotator and broader category generalization were not tested; the paper has no formal limitations section.
  • Evident: the ensemble approach did not improve nuclear segmentation (Extended Data Fig. 2), and the ensemble benefit depended on the Cellpose architecture broadcasting the style vector (Extended Data Fig. 1).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41592-022-01663-4

Cites

54 of 54 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

17 of 21 cited sources checked · 6 supported · 0 flagged · accuracy not yet measured

  • Findingown data

    The style vectors for all images clustered into nine different classes using the Leiden algorithm.

    “We took the style vectors for all images and clustered them into nine different classes using the Leiden algorithm, illustrated on a t-SNE (t-distributed stochastic neighbor embedding) plot in Fig. 2a (refs. 25,26).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [25] Traag, V. A., Waltman, L. & Van Eck, NeesJan From Louvain to Leiden: guaranteeing well-connected communities. Sci. Rep. 9, 5233 (2019). source
    • cites [26] Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 9, 2579–2605 (2008).

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Findingcitation

    The Cellpose models outperform the LiveCell model released with the LiveCell dataset.

    “The Cellpose models also outperform the LiveCell model released with the LiveCell dataset31.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [31] Lee, Y. & Park, J. Centermask: real-time anchor-free instance segmentation. In Proc. of the IEEE/CVF conference on computer vision and pattern recognition 13906–13915 (IEEE, 2020). source

      The Introduction and §3.2 compare CenterMask with other instance-segmentation methods on COCO, not Cellpose with a LiveCell model. Searches for Cellpose, LiveCell, and cell-segmentation comparisons found no passage in our copy addressing the claimed result. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
  • Findingown data

    The ensemble of models for nuclear segmentation did not show an improvement compared to the generalist model.

    “However, we did not see an improvement for the ensemble of models compared to the generalist model (Extended Data Fig. 2), consistent with the results of ref. 27.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [27] Caicedo, J. C. et al. Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl. Nat. Methods 16, 1247–1253 (2019). source

      The Results section (‘Top solutions improve usability and accuracy of nucleus segmentation’; page unavailable) reports that five models trained separately for image types did not reach the performance of the top solution. That is consistent with the citing sentence’s general point, but it is not a direct comparison of an ensemble against a generalist model: the paper describes the top solution itself as a 32-network ensemble. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
  • Interpretationreasoning

    The within-human upper bound is by construction higher than any across-human upper bound because it excludes inconsistencies in segmentation styles between different annotators.

    “Note that the within-human upper bound is by construction higher than any ‘across-human’ upper bound6, because it excludes inconsistencies in segmentation styles between different annotators.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [6] Greenwald, N. F. et al. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nat. Biotechnol. 40, 555–565 (2021). source

      The section ‘Mesmer achieves human-level performance for segmentation’ says expert annotators can disagree and reports comparisons between different annotators. The paper does not establish a within-human upper bound or show that it must be higher than any across-human upper bound because of differences in segmentation style. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
  • Findingreasoning

    The approach requires a user to segment 500-1,000 ROI offline or 100-200 ROI with a human-in-the-loop approach.

    “We demonstrate that this approach is: (1) necessary, because annotation styles can vary dramatically between different annotators; (2) efficient, because it only requires a user to segment 500–1,000 ROI offline or 100–200 ROI with a human-in-the-loop approach and (3) effective, because models created this way have similar accuracy to human experts.”
    Quote in the paperSubstantial claim
  • May combine two claims. 32
  • Classified as probably not a substantial claim. 17

  • May combine two claims.

    Premisecitation

    The human-in-the-loop approach has been demonstrated for simple ROI such as nuclei and round cells, but not for cells with complex morphologies that require full cytoplasmic segmentation.

    “This approach has been demonstrated for simple ROI such as nuclei and round cells, which allow for weak annotations such as clicks and squiggles18,19, but not for cells with complex morphologies that require full cytoplasmic segmentation.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [18] Alemi Koohbanani, N., Jahanifar, M., Zamani Tajadin, N. & Rajpoot, N. Nuclick: a deep learning framework for interactive segmentation of microscopic images. Med. Image Anal. 65, 101771 (2020). source

      The Abstract and §2.3 describe interactive segmentation of nuclei and cells using clicks and glands using squiggles. The paper does not establish the claimed limitation for cells with complex morphologies requiring full cytoplasmic segmentation; it also reports segmentation of whole white blood cells (§4.5). (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [19] Sugawara, K., Çevrim, C. ¸ & Averof, M. Tracking cell lineages in 3D by incremental deep learning. eLife 11, e69380 (2022). source

      The Results and discussion section describes human-in-the-loop detection and tracking of nuclei using sparse ellipsoid annotations and user proofreading. It does not demonstrate round-cell annotation with clicks or squiggles, nor establish the broader claim that this approach has not been demonstrated for complex cell morphologies requiring full cytoplasmic segmentation. No page number is available in our copy. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The human-in-the-loop strategy used to annotate the TissueNet dataset took two human years of crowdsourced work for 14 image categories.

    “This is the strategy used to annotate the TissueNet dataset, which in total took two human years of crowdsourced work for 14 image categories6,17.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [6] Greenwald, N. F. et al. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nat. Biotechnol. 40, 555–565 (2021). source

      The ‘A human-in-the-loop approach for constructing TissueNet’ section describes crowdsourced annotation and reports more than 4,000 person-hours, equivalent to nearly two person-years. It describes data from nine organs and six imaging platforms, but does not support the claim of 14 image categories. (LLM, conf=0.90)

      Experimental — accuracy not yet measured
    • cites [17] Gurari, D. et al. How to collect segmentations for biomedical images? A benchmark evaluating the performance of experts, crowdsourced non-experts, and algorithms. In Proc. 2015 IEEE Winter Conference on Applications of Computer Vision 1169–1176 (IEEE, 2015). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Using an iterative approach, a 3D dataset of nuclei was segmented in approximately one month.

    “For example, using an iterative approach19, a 3D dataset of nuclei was segmented in approximately one month.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [19] Sugawara, K., Çevrim, C. ¸ & Averof, M. Tracking cell lineages in 3D by incremental deep learning. eLife 11, e69380 (2022). source

      The Results and discussion report that an interactive ELEPHANT workflow completed cell-lineage tracking in approximately one month, and Materials and methods describe iterative annotation of nuclei in the 3D dataset. The month-long result was reconstruction and validation of cell lineages, not specifically segmentation of a dataset of nuclei. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The pretraining approach has been successful for various machine vision problems.

    “The pretraining approach has been successful for various machine vision problems28–30.”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [28] Zamir, A. R. et al. Taskonomy: disentangling task transfer learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition 3712–3722 (IEEE, 2018). source
    • cites [29] Da Silva, F. L. & Costa, A. H. R. A survey on transfer learning for multiagent reinforcement learning systems. J. Artif. Int. Res. 64, 645–703 (2019). source

      Section 5.8 describes pre-training a deep network from demonstrations for reinforcement learning, but does not address success across machine-vision problems. Section 6.3 discusses image-derived features in video games, not the success of pretraining for machine vision. Searches for pre-training in machine vision and pretrained networks in computer vision found no passage establishing the claim; page numbers were unavailable in our copy. (LLM, conf=0.74)

      Experimental — accuracy not yet measured
    • cites [30] Morid, M. A., Borjali, A. & Del Fiol, G. A scoping review of transfer learning research on medical image analysis using imagenet. Comput. Biol. Med. 128, 104115 (2021). source

  • May combine two claims.

    Premisecitation

    The LiveCell dataset consists of 3,188 training and 1,516 test images of eight different cell lines collected using phase-contrast microscopy, and each image has manual segmentations of the cells.

    “The LiveCell dataset consists of 3,188 training and 1,516 test images of eight different cell lines collected using phase-contrast microscopy, and each image has manual segmentations of the cells (https://sartorius-research.github.io/LIVECell/)7.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [7] Edlund, C. et al. Livecell-a large-scale dataset for label-free live cell segmentation. Nat. Methods 18, 1038–1045 (2021). source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The TissueNet dataset consists of 2,601 training and 1,249 test images of six different tissue types collected using fluorescent microscopy on six different platforms, and each image has manual segmentations of the cells and the nuclei.

    “The TissueNet dataset consists of 2,601 training and 1,249 test images of six different tissue types collected using fluorescent microscopy on six different platforms, and each image has manual segmentations of the cells and the nuclei (https://datasets.deepcell.org/)6.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [6] Greenwald, N. F. et al. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nat. Biotechnol. 40, 555–565 (2021). source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The LiveCell study used a different average precision score from ours, which additionally requires a confidence score per ROI, while we used the average precision formulation from the Data Science Bowl challenge and other studies.

    “Finally, we point out that the LiveCell study used a different average precision score from ours, which additionally requires a confidence score per ROI, while we used the average precision formulation from the Data Science Bowl challenge and other studies12,27,35.”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [12] Hollandi, R. et al. nucleaizer: a parameter-free deep learning framework for nucleus segmentation using image style transfer. Cell Systems 10, 453–458 (2020). source
    • cites [27] Caicedo, J. C. et al. Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl. Nat. Methods 16, 1247–1253 (2019). source
    • cites [35] Schmidt, U., Weigert, M., Broaddus, C. & Myers, G. Cell detection with star-convex polygons. In Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention 265–273 (Springer, 2018). source

  • Classified as probably not a substantial claim.

    Premisecitation

    Large amounts of data are necessary to train powerful models in general computer vision.

    “Our conclusions may seem at odds with the general intuition from the computer vision literature, where large amounts of data are necessary to train powerful models33,34.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [33] Russakovsky, O. et al. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115, 211–252 (2015). source
    • cites [34] Lin, T. Y. et al. Microsoft COCO: common objects in context. In Proc. European Conference on Computer Vision 740–755 (Springer, 2014). source

      The Introduction (p. 1) and Related Work (p. 3) say that ImageNet’s millions of images enabled breakthroughs in object classification and detection. That supports the broad intuition that large datasets can help produce powerful vision models, but the paper does not establish that large amounts of data are necessary to train such models in computer vision generally. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    The LC2 model ignored ROI in very dense regions, unlike the LC1 model that segmented everything.

    “the LC2 model ignored ROI in very dense regions, unlike the LC1 model that segmented everything (Fig. 2d(v))”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given

    The human annotator observed that the model's mistakes were becoming more subjective, and were often due to uncertain cues in the image.

    “Qualitatively, the human annotator observed that the model’s mistakes were becoming more subjective, and were often due to uncertain cues in the image.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Nuclei were not segmented in the Cellpose dataset if they were missing a cytoplasm or membrane label, but they were always labeled in the TissueNet dataset.

    “For example, nuclei were not segmented in the Cellpose dataset if they were missing a cytoplasm or membrane label (Fig. 1a(i)), but they were always labeled in the TissueNet dataset (Fig. 1b(i),(iii)).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Performance improves dramatically with 504 training ROI, and then improves much more slowly until it reaches the maximum at 81,832 ROI.

    “Performance improves dramatically with 504 training ROI (equivalent to two training images), and then improves much more slowly until it reaches the maximum at 81,832 ROI.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The TN1 model drew smaller regions around each nucleus than the TN2 model, which extended the ROI until they touched each other.

    “the TN1 model drew smaller regions around each nucleus than the TN2 model, which extended the ROI until they touched each other (Fig. 2d(i))”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the next image, only 32 new ROI had to be manually added, which continued to decrease on the third, fourth and fifth images.

    “On the next image, only 32 new ROI had to be manually added, which continued to decrease on the third, fourth and fifth images.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The LC1 model overall identified more cells than the LC4 model, which specifically ignored larger ROI.

    “the LC1 model overall identified more cells than the LC4 model, which specifically ignored larger ROI (Fig. 2d(iv))”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Pretrained Cellpose models improved faster than models trained from scratch: the pretrained model reaches an average precision of 0.73 at 426 training ROI versus 0.68 for the model trained from scratch.

    “Furthermore, pretrained Cellpose models improved faster than the models trained from scratch: the pretrained model reaches an average precision of 0.73 at 426 training ROI versus 0.68 average precision for the model trained from scratch.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Some TissueNet categories included the entire cytoplasm, others excluded portions of the cytoplasm, and some focused exclusively on the nucleus.

    “Some TissueNet categories also included the entire cytoplasm (Fig. 1b(i)), but others excluded portions of the cytoplasm (Fig. 1b(ii),(iii)) or even focused exclusively on the nucleus (Fig. 1b(iv)).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    There were four image classes composed mainly of fluorescent cell images, four classes composed mainly of phase-contrast images, and a ninth class including a wide variety of images.

    “There were four image classes composed mainly of fluorescent cell images (CP, TN1, TN2, TN3), four classes composed mainly of phase-contrast images (LC1, LC2, LC3, LC4) and a ninth class including a wide variety of images (CPx) (Fig. 2b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Diffuse processes were not segmented in the Cellpose dataset but were always segmented in the LiveCell dataset.

    “Processes that were diffuse were not segmented in the Cellpose dataset (Fig. 1a(ii)) but they were always segmented in the LiveCell dataset (Fig. 1c(iv)).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Areas of high density and low confidence were given annotations in the Cellpose dataset and in some LiveCell categories, while they were often not segmented in other LiveCell categories.

    “Finally, areas of high density and low confidence were nonetheless given annotations in the Cellpose dataset and in some LiveCell categories (Fig. 1a(iv),c(i)), while they were often not segmented in other LiveCell categories (Fig. 1c(ii)–(iv)).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The CPx model segmented everything that looked like an object, while the TN1 model selectively identified only bright objects, assigning dim objects to the background.

    “the CPx model segmented everything that looked like an object, while the TN1 model selectively identified only bright objects, assigning dim objects to the background (Fig. 2d(iii))”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    All image classes had improvements in the range of 0.01-0.06 for the average precision score, with the largest improvements at higher IoU thresholds and for the most diverse image class (CPx).

    “All image classes had improvements in the range of 0.01–0.06 for the average precision score, with the largest improvements observed at higher intersection-over-union (IoU) thresholds, and for the most diverse image class (CPx).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Annotation styles can vary dramatically between different annotators.

    “We demonstrate that this approach is: (1) necessary, because annotation styles can vary dramatically between different annotators; (2) efficient, because it only requires a user to segment 500–1,000 ROI offline or 100–200 ROI with a human-in-the-loop approach and (3) effective, because models created this way have similar accuracy to human experts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The CP model carefully tracked the precise edges of cells while the TN3 model ignored processes.

    “the CP model carefully tracked the precise edges of cells while the TN3 model ignored processes (Fig. 2d(ii))”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    For the TissueNet category, many cells were correctly segmented by the pretrained Cellpose model, but nuclei without cytoplasm were always ignored.

    “For the TissueNet category, the human annotator observed that many cells were correctly segmented by the pretrained Cellpose model, but nuclei without cytoplasm were always ignored, which is likely due to the segmentation style used in the original Cellpose dataset (Fig. 1a(i)).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The rapid initial improvement is also seen on average for multiple models trained with different subsets of the data and on all TissueNet categories.

    “The rapid initial improvement is also seen on average for multiple models trained with different subsets of the data and on all TissueNet categories (Fig. 3b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    We observed slight performance improvements for the models trained from scratch but not from pretrained for 300 epochs of training compared to 100 epochs.

    “We observed slight performance improvements for the models trained from scratch but not from pretrained for 300 epochs of training compared to 100 epochs.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The ground-truth segmentations in the TissueNet and LiveCell datasets are accurate.

  • May combine two claims.

    Findingown data

    For the LiveCell dataset, 52 out of 127 ROI had to be drawn manually on the first image, but only 18 out of 293 ROI had to be drawn on the fifth image.

    “A similar process was observed for images from the LiveCell dataset (Fig. 4c), where 52 out of 127 ROI had to be drawn manually on the first image, but only 18 out of 293 ROI had to be drawn on the fifth image.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In both cases, 100-200 manually segmented ROI were sufficient to achieve near-maximal accuracy and the process only required 1-2 hours of the user's time.

    “In both cases, 100–200 manually segmented ROI were sufficient to achieve near-maximal accuracy and the process only required 1–2 hours of the user’s time.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    The diversity of the Cellpose training set allows the pretrained Cellpose model to generalize well to new images and provides a good starting set of parameters for further fine-tuning on new image categories.

    “The diversity of the Cellpose training set allows the pretrained Cellpose model to generalize well to new images, and provides a good starting set of parameters for further fine-tuning on new image categories.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    None of these differences are mistakes; they are different styles of segmenting the same images.

    “None of these differences are mistakes. Instead, they are different styles of segmenting the same images, each of which may be preferred by a user depending on circumstances.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    The Cellpose dataset is smaller but more diverse than the TissueNet and LiveCell datasets.

    “The Cellpose dataset is smaller but more diverse than the TissueNet and LiveCell datasets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    This increase in performance happens despite the Mesmer model being trained with up to 200,000 training ROI from each image category, and is likely explained by differences between the architecture of the segmentation models.

    “This increase in performance happens despite the Mesmer model being trained with up to 200,000 training ROI from each image category, and is likely explained by differences between the architecture of the segmentation models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The original Cellpose model may generalize across varying image types, but cannot generalize across different segmentation styles.

    “This suggests that the original Cellpose model may generalize across varying image types, but cannot generalize across different segmentation styles.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    The TissueNet dataset contained 13 image categories with at least ten training images each, and the LiveCell dataset contained eight.

    “The TissueNet dataset contained 13 image categories with at least ten training images each, and the LiveCell dataset contained eight.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Retraining times will vary relatively little with the type of GPU used because batch sizes are small (eight).

    “We expect similar run time performance for other GPUs and we expect that retraining times will vary relatively little with the type of GPU used because our batch sizes are small (eight).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given

    It is possible, although not desirable, to run the human-in-the-loop process on the CPU, where retraining times of at least several minutes should be expected.

    “It is possible, although not desirable, to run the human-in-the-loop process on the CPU, where retraining times of at least several minutes should be expected.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Similar run time performance is expected for other GPUs.

    “We expect similar run time performance for other GPUs and we expect that retraining times will vary relatively little with the type of GPU used because our batch sizes are small (eight).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Cellpose default model ('cyto') was trained on 540 images of cells and objects with one or two channels (if the image had a nuclear channel).

    “The Cellpose default model (‘cyto’) was trained on 540 images of cells and objects with one or two channels (if the image had a nuclear channel).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5] Stringer, C., Wang, T., Michaelos, M. & Pachitariu, M. Cellpose: a generalist algorithm for cellular segmentation. Nat. Methods 18, 100–106 (2021). source

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The discrepancy between our conclusions and general computer vision intuition may be due to differences of scope between cell segmentation and general computer vision tasks.

    “The discrepancy may be due to differences of scope between cell segmentation and general computer vision tasks.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Deep learning models for general computer vision tasks need to perform well on a large diversity of test images, and therefore require a large diversity of training images.

    “Deep learning models for general computer vision tasks need to perform well on a large diversity of test images, and therefore require a large diversity of training images.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    More varied training data is now easy to generate with the human-in-the-loop approach from Cellpose 2.0.

    “We renew our calls for the community to contribute more varied training data, which is now easy to generate with the human-in-the-loop approach from Cellpose 2.0.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisenone given

    The LiveCell authors concluded that segmentation performance continues to increase when adding more training data.

    “The LiveCell authors concluded that segmentation performance continues to increase when adding more training data.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A biologist may prefer one segmentation style over another depending on the purpose of their study.

    “This variability hints at a fundamental aspect of biological segmentation: there are often multiple correct solutions, and a biologist may prefer one segmentation style over another depending on the purpose of their study.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    The TissueNet authors concluded that performance saturates at 104-105 training ROI.

    “The TissueNet authors concluded that performance saturates at 104–105 training ROI.”
    Quote in the paperSubstantial claim
45 of 45

  • The paper relies on this without support.

    Pretrained neural network models for biological segmentation can provide good out-of-the-box results for many image types.

    “Pretrained neural network models for biological segmentation can provide good out-of-the-box results for many image types.”

  • The paper relies on this without support.

    Pretrained models do not allow users to adapt the segmentation style to their specific needs and can perform suboptimally for test images that are very different from the training images.

    “However, such models do not allow users to adapt the segmentation style to their specific needs and can perform suboptimally for test images that are very different from the training images.”

  • The paper relies on this without support.

    Neural networks struggle to generalize to out-of-distribution data, that is new images that look fundamentally different from anything seen during training.

    “However, neural networks struggle to generalize to out-of-distribution data, that is new images that look fundamentally different from anything seen during training.”

  • The paper relies on this without support.

    The diversity of biological image acquisition methods poses a grand challenge to automated segmentation approaches, which have traditionally been developed for specific applications, and fail when applied to new types of data.

    “This diversity of methods poses a grand challenge to automated segmentation approaches, which have traditionally been developed for specific applications, and fail when applied to new types of data.”

  • The paper relies on this without support.

    The Cellpose dataset is smaller but more diverse than the TissueNet and LiveCell datasets.

    “The Cellpose dataset is smaller but more diverse than the TissueNet and LiveCell datasets.”

  • The paper relies on this without support.

    The ground-truth segmentations in the TissueNet and LiveCell datasets are accurate.

    “each image has manual segmentations of the cells and the nuclei”
    “each image has manual segmentations of the cells”

  • The paper relies on this without support.

    The test images in the TissueNet and LiveCell datasets are representative of the image classes they are meant to evaluate.

    “The TissueNet dataset consists of 2,601 training and 1,249 test images of six different tissue types”
    “The LiveCell dataset consists of 3,188 training and 1,516 test images of eight different cell lines”

  • The paper relies on this without support.

    The Cellpose model cannot predict overlapping ROI.

    “The Cellpose model cannot predict overlapping ROI, therefore the overlapping pixels were reassigned to the mask with the closest centroid.”
8 of 8

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • LiveCell models are attributed to TissueNet training images

    Minor

    The Fig. 5 caption says the A172 LiveCell models were trained on the images in Fig. 4b, but Fig. 4b shows TissueNet images. Fig. 4c shows the A172 LiveCell training images.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. Summerfield, M. Rapid GUI Programming with Python and Qt: The Definitive Guide to PyQt Programming (Pearson Education, 2007).
  2. Kluyver, T. et al. Jupyter notebooks—a publishing format for reproducible computationalworkflows. In Proc. 20th International Conference on Electronic Publishing: Positioning and Power in Academic Publishing: Players, Agents and Agendas (eds Loizides, F. & Schmidt, B.) 87–90 (IOS Press, 2016).

Cited by

Sorted by citations

Related

Authors
Pachitariu, M., & Stringer, C.
Affiliations
Howard Hughes Medical Institute (US)
Corresponding author
  • Pachitariu, Marius
  • Stringer, Carsen
ORCID
Venue
Nature Methods
Publisher
Springer Science and Business Media LLC
Published
2022-11-07
Volume
19
Issue
12
Pages
1634-1641
ISSN
  • 1548-7091 (ISSN-L)
  • 1548-7105
DOI
10.1038/s41592-022-01663-4
PMC
PMC9718665
OpenAlex
W4308361273
Semantic Scholar
0980e47ee671041f51460c2371543b2b2de7d071
Cited by
1,381
References
54
Citations by year
2026: 359 · 2025: 481 · 2024: 352 · 2023: 168 · 2022: 19 · 2020: 1 · 2019: 1
FWCI
107.92
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Methods (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
Funders
Howard Hughes Medical Institute
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Biophysics › Cell Image Analysis Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
Keywords
cell segmentation, fine-tuning, human-in-the-loop annotation, model zoo, ensemble learning
MeSH terms
Humans, Image Processing, Computer-Assisted, Software, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-20.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-20.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review