Pith. sign in

REVIEW 3 major objections 6 minor 62 references

A hierarchy-aware SSL method keeps fine cell morphology from being buried by imaging modality.

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 ↗

T0 review · grok-4.5

2026-07-11 19:51 UTC pith:DVN33STV

load-bearing objection Solid empirical SSL method for single-cell microscopy: segmentation teacher + stability-weighted HDBSCAN prototypes give real gains on a large curated corpus and a useful drug-perturbation task, even if the “true hierarchy” story is only partly validated. the 3 major comments →

arxiv 2607.04353 v1 pith:DVN33STV submitted 2026-07-05 cs.CV cs.AI

HASSL: Hierarchy-Aware Self-Supervised Learning Framework for Single Cell Microscopy

Julius Riel , Vishwa Mohan Singh , Sai Anirudh Aryasomayajula , Anuun Chinbat , Hannes Leonhard , Moritz Ladenburger , Frederik Alexander , Vishisht Choudhary
show 5 more authors
Fabio Laredo Giacomo Masserdotti Thorben Prein Carsten Marr Amirhossein Kardoost
This is my paper
classification cs.CV cs.AI
keywords self-supervised learninghierarchical representationsingle-cell microscopyHDBSCANdistillationmorphological profilingcell embedding
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.

Self-supervised models trained on microscopy images often let coarse factors such as imaging modality dominate the embedding space, so cells of different types that look similar under the same setup collapse together and the same type imaged differently is split apart. This paper claims that two complementary pieces of supervision can reverse that bias: a second distillation teacher that sees zero-shot segmentation masks, and a contrastive term that builds hierarchical prototypes with HDBSCAN and pulls each cell toward its own ancestors while repelling only its siblings, weighted by cluster stability. Trained on a new 2.3-million-cell multi-modality corpus spanning 208 classes, the resulting embeddings improve retrieval, clustering agreement, and especially a biologically relevant drug-identification task that hinges on subtle morphological change. A reader who cares about single-cell phenotyping or mechanism-of-action inference therefore has a concrete recipe for representations that respect multi-level biological structure rather than acquisition artefacts.

Core claim

Adding a segmentation-guided second teacher to DINO-style distillation together with a stability-weighted HDBSCAN hierarchy-aware contrastive loss produces single-cell embeddings whose nearest-neighbour structure and downstream classifiers better track true morphological subtypes instead of modality superclusters, raising average top-K accuracy by 2.8 percent, top-9 retrieval on the deepest-hierarchy subset by 6.3 percent, and weighted F1 on drug classification from perturbed morphology by 7.8 percent.

What carries the argument

The double-teacher objective (image teacher plus segmentation teacher) plus the HDBSCAN prototype loss that mines ancestor positives and sibling negatives, each re-weighted by cluster stability λ, so the hinge enforces hierarchical separation without labels.

Load-bearing premise

That zero-shot segmentation masks are faithful enough morphology priors, and that the condensed trees HDBSCAN builds on each batch of embeddings recover the true multi-level biological hierarchy well enough for the mined prototypes to sharpen the right boundaries.

What would settle it

Re-train the identical pipeline after replacing CellposeSAM masks with pure noise or after replacing HDBSCAN with flat random clusters of the same sizes; if the reported gains in multi-level retrieval and drug-identification F1 disappear, the hierarchy-aware claim fails.

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

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 HASSL, a self-supervised framework for single-cell microscopy that aims to preserve multi-level biological structure rather than letting coarse factors (especially imaging modality) dominate the latent space. It combines (i) a double-teacher DINO-style distillation objective in which a segmentation teacher, driven by zero-shot CellposeSAM masks, supplies morphology-aware targets (Eqs. 1–6), and (ii) a stability-λ-weighted hierarchical contrastive loss that builds in-batch HDBSCAN condensed trees, mines ancestor positives and sibling negatives, and applies a hinge objective (Eqs. 7–16, Alg. 2). The method is trained on a curated 2.3M single-cell crop corpus from 20 datasets (208 classes) and evaluated with k-NN retrieval, clustering agreement, frozen-MLP cell-type classification, held-out HPA transfer, and drug identification from Allen Institute perturbation images. Reported gains over DINOv3 and cell-specific SSL baselines include ~+2.8% average top-K accuracy, +6.3% top-9 on multi-level-hierarchy subsets, and +7.8% weighted F1 on the drug task.

Significance. Hierarchy suppression by modality and batch effects is a genuine, practically important failure mode of SSL on cellular images. The paper contributes a usable drop-in objective, a large multi-modality single-cell benchmark with public code and Hugging Face data, and consistent empirical improvements across retrieval, clustering, and two external downstream settings (AICS drug ID and HPA). Ablations (Table 1) and the depth-split analysis (Fig. 4) give some support that both components matter and that gains concentrate on deeper hierarchies. If the hierarchy-aware interpretation holds, the work is a useful step toward morphology-centric cellular foundation models; even under a weaker “better morphology features” reading the engineering and benchmark contributions remain valuable for the community.

major comments (3)
  1. [Sec. 3.2 / Eqs. 8–16] Abstract, Sec. 1, and Sec. 3.2 (Eqs. 8–16, Alg. 2): The central claim is not merely improved morphology features but that the method recovers and sharpens biologically meaningful hierarchical substructure. The load-bearing assumption is that in-batch HDBSCAN condensed trees (min_cluster_size=2) and sibling/ancestor prototype mining align with true multi-level biology rather than modality, batch, or density artifacts of the current embedding. Ablations show flat DBSCAN and unweighted HDBSCAN underperform, and Fig. 4 shows larger gains on depth>1 subsets, but there is no quantitative check (e.g., AMI/NMI of induced tree levels vs. known subtype/parent labels, or purity of mined siblings against biological taxonomies). Without such a check, the hierarchy-aware interpretation remains under-supported relative to a simpler morphology-prior reading. A modest validation on labeled hierarchical s
  2. [Sec. 3.1] Sec. 3.1 and training methodology: Zero-shot CellposeSAM masks are treated as a sufficiently faithful, label-free morphology prior that initiates break-up of modality superclusters. The paper correctly notes that masks can tolerate noise, but there is no ablation on segmentation quality (e.g., degraded masks, alternative segmenters, or mask-free control beyond the “without Double Teacher” row). Because the double-teacher term is one of the two named contributions and is annealed via γ, a short sensitivity or quality-robustness experiment is needed to show that gains are not contingent on unusually clean CellposeSAM outputs on this particular corpus.
  3. [Table 1 / Sec. 4.2] Table 1 and Sec. 4.2: Free parameters (γ ramp 0→0.2, HDBSCAN weight ramp in the last 20 epochs, hinge margin m, min_cluster_size=2, Sinkhorn temperature, ε) are schedule-dependent and only partially ablated. The unweighted-HDBSCAN and DBSCAN rows help, but there is no sensitivity sweep on min_cluster_size or on when/how strongly the hierarchical term is introduced. Given that min_cluster_size=2 maximizes depth and can produce noisy leaves, a brief stability analysis (or reporting variance over a few schedule choices) would make the gains more credible as method properties rather than schedule-tuned effects.
minor comments (6)
  1. [Fig. 3] Fig. 3 caption and main text: “Florucent” appears to be a typo for “Fluorescent.”
  2. [Table 3] Table 3 lists “HASSL (w/o DT)” twice with different numbers; the second occurrence is likely meant to be “HASSL (w/o HDBSCAN)” or similar—please correct the row labels.
  3. [Sec. 3.2] Sec. 3.2: The notation for the sibling set and path (Sib(c), π(c), Ch(·), ci,k vs cik) is slightly inconsistent across equations; a single consistent indexing convention would help.
  4. [Sec. 2] Related Work: HCSC is correctly positioned as closest prior work; a one-sentence clarification of how stability-weighted sibling mining differs from HCSC’s recursive k-NN positive expansion would help readers who know that line of work.
  5. [Sec. 4.1 / Fig. 4] Supplementary Table 5 and Fig. 5 are useful; consider briefly stating in the main text how many of the 20 datasets have depth>1 so that Fig. 4’s split is easier to interpret without the supplement.
  6. [Front matter] Version note and arXiv header: the preprint framing is fine; ensure final camera-ready removes the “pre-peer-review” banner and any placeholder dates if present.

Circularity Check

0 steps flagged

No circularity: SSL losses and HDBSCAN prototypes are defined label-free from embeddings; reported retrieval/F1 gains are measured on held-out external labels and datasets, not forced by construction.

full rationale

This is a standard empirical self-supervised learning paper. The double-teacher distillation (Eqs. 1–6, Alg. 1) uses zero-shot CellposeSAM masks as a weak prior and DINO-style Sinkhorn targets; the hierarchy-aware term (Eqs. 7–16, Alg. 2) runs in-batch HDBSCAN (min_cluster_size=2) on current student embeddings to mine ancestor/sibling prototypes and applies a stability-weighted hinge. Neither objective is defined in terms of the evaluation labels (cell classes, drug IDs). Training is on the curated 2.3 M corpus; metrics (Tables 1–4, Figs. 3–4) are k-NN retrieval, NMI/AMI, and MLP classification on held-out splits plus completely unseen AICS perturbation and HPA sets. Ablations isolate components without circular reduction. No parameter is fitted to a target quantity and then re-reported as a prediction; no uniqueness theorem or load-bearing self-citation forces the result; no known empirical pattern is merely renamed. The design choices (annealed γ, λ weights, min_cluster_size) are ordinary hyperparameters, not definitional equivalences. The paper is therefore self-contained against its external benchmarks; circularity score is zero.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central claim is empirical: the proposed losses improve morphology-aware hierarchical structure in embeddings. It rests on standard SSL math, domain assumptions that masks and HDBSCAN trees are useful proxies for morphology/hierarchy, and several annealed hyperparameters chosen by the authors. No new physical entities; the 'invented' pieces are the training components themselves.

free parameters (6)
  • double-teacher mix weight γ = 0 → 0.2
    Linearly ramped from 0 to 0.2 over pre-training; controls how much segmentation targets influence the student. Chosen by schedule, not derived.
  • HDBSCAN loss weight = 0 → 0.1
    Ramped from 0 to 0.1 in the final 20 epochs; directly scales the hierarchy-aware term that drives the headline gains.
  • hinge margin m
    Margin in the stability-weighted hinge contrastive loss (Eq. 15); free design choice affecting separation strength.
  • HDBSCAN min_cluster_size = 2
    Set to 2 to maximize tree depth; strongly shapes which prototypes exist and thus the contrastive pairs.
  • Sinkhorn-Knopp temperature Temp
    Teacher target sharpening temperature following DINOv3; affects distillation targets.
  • stability transform epsilon ε = small ε > 0
    Clamping constant in ϕ_ε for positive/negative λ weights; numerical/design choice.
axioms (5)
  • domain assumption Zero-shot CellposeSAM (or any generalist) segmentation masks provide a weak but useful morphology prior without label leakage.
    Sec. 3.1 bases the second teacher entirely on these masks; if masks are systematically wrong, morphology breakout fails.
  • domain assumption In-batch HDBSCAN condensed trees on L2-normalized embeddings approximate the multi-level semantic hierarchy of cell types well enough for ancestor/sibling prototype mining.
    Sec. 3.2; the hierarchy-aware loss is defined from this tree, not from ground-truth taxonomy.
  • domain assumption DINO-style multi-crop self-distillation with Sinkhorn-Knopp centering is a valid base objective for single-cell crops.
    Method is built as an extension of DINOv3 defaults (Sec. 3.1, 4.2).
  • standard math Cosine similarity on unit-normalized embeddings plus a hinge margin is an appropriate geometry for hierarchical prototype attraction/repulsion.
    Eqs. 7–16; standard contrastive geometry, not proved optimal for cells.
  • ad hoc to paper Aggregating 20 heterogeneous public datasets into one 2.3M single-cell crop corpus with 90/10 within-dataset splits yields a fair multi-modality benchmark.
    Sec. 4.1; curation choices (pseudo-labels for unlabeled sets, oriented-box crops) define the evaluation surface.
invented entities (2)
  • Double-teacher DINO distillation (image EMA teacher + segmentation teacher with L_img→img, L_seg→seg, L_img→seg) no independent evidence
    purpose: Bias embeddings toward morphology and begin breaking modality superclusters without class labels.
    Introduced in Sec. 3.1 as a paper-specific training architecture; no independent external validation beyond this work's ablations.
  • Stability-λ-weighted hierarchical prototype hinge loss with sibling negatives from HDBSCAN MST no independent evidence
    purpose: Sharpen decision boundaries between related subtypes at multiple hierarchy levels using label-free prototypes.
    Sec. 3.2; related to HCSC-style hierarchical prototypes but with HDBSCAN stability weighting and sibling mining specific to this paper.

pith-pipeline@v1.1.0-grok45 · 24515 in / 3804 out tokens · 44199 ms · 2026-07-11T19:51:10.295297+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HASSL: Hierarchy-Aware Self-Supervised Learning Framework for Single Cell Microscopy." pith.science (2026). https://pith.science/paper/DVN33STV

@misc{pith2026260704353,
  author       = {Pith},
  title        = {Pith review of: HASSL: Hierarchy-Aware Self-Supervised Learning Framework for Single Cell Microscopy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVN33STV}},
  note         = {Machine review of arXiv:2607.04353}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Hierarchical structure is common in image data, where fine-grained clusters often merge into larger, coarser semantic groups. In biological cell images, current self-supervised learning models often suppress this hierarchy, as coarse factors such as imaging modality can obscure finer morphological attributes in the latent space. We propose a hierarchy-aware self-supervised training framework to address this problem. Our method combines two components: a distillation framework with a segmentation teacher to improve morphological awareness in the latent space, and a hierarchy-aware contrastive loss based on HDBSCAN to improve decision boundaries between closely related subtypes at different hierarchical levels. Together, these components reduce the tendency of self-supervised learning to overemphasize coarse factors and instead align embeddings with semantic and morphological cues. This yields biologically meaningful sub-clusters driven by fine morphological detail. We train and evaluate our method on a curated corpus of 2.3 million single cells aggregated from 20 microscopy datasets, both labeled and unlabeled, covering 208 cell classes. Our method improves over baseline and counterpart methods, increasing average top-K accuracy by 2.8%, top-9 retrieval on the dataset with the deepest hierarchy by 6.3%, and downstream F1-score for biologically relevant drug classification from perturbed cell morphology by 7.8%.

Figures

Figures reproduced from arXiv: 2607.04353 by Amirhossein Kardoost, Anuun Chinbat, Carsten Marr, Fabio Laredo, Frederik Alexander, Giacomo Masserdotti, Hannes Leonhard, Julius Riel, Moritz Ladenburger, Sai Anirudh Aryasomayajula, Thorben Prein, Vishisht Choudhary, Vishwa Mohan Singh.

Figure 1
Figure 1. Figure 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our hierarchy-aware representation learning pipeline. (1) Segmentation-map generation: microscopy images are zero-shot segmented [49], which is used as a proxy for morphological features (2) Double-Teacher distillation: a student ViT is trained with EMA from a global image teacher and a segmentation teacher, pro￾moting subcluster breakout in the latent space. (3) Hierarchical cluster reinforcem… view at source ↗
Figure 3
Figure 3. Figure 3: Latent space visualization from our model on the test set, produced using t￾SNE [36] with dimension=2. In the internal stem cell data (orange), the double teacher distillation brings superclusters of different modalities but the same cell type closer, and HDBSCAN still keeps the subclusters compact and discernible. In the global space, the Florucent Cluster is the 2nd closest neighbour to the Brightfield c… view at source ↗
Figure 4
Figure 4. Figure 4: Top-k retrieval accuracies for subsets containing single-level hierarchy datasets (left, depth=1) and multi-level hierarchy datasets (right, depth¿1). Here, HASSL is the best performing model, with an improvement of 6.3% on multi-level hierarchy datasets [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Radial Diagram showing the modalities and cell classes available in our dataset. The dataset exhibits multi-level hierarchies, including modalities, dataset/collection method, cell types, and subtypes. D Result Comparison The top-K accuracy for each dataset is shown in [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of the top-K retrieval results on all the individual datasets in our collection. Our model consistently returns high scores and beats the baselines in the majority of the cases, leading to a high average accuracy score across the test dataset [PITH_FULL_IMAGE:figures/full_fig_p025_6.png] 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

62 extracted references · 7 canonical work pages

  1. [1]

    [dataset] (2018), https://www.allencell.org/drug-perturbation-pilot.html, accessed 2026-02- 24

    Allen Institute for Cell Science: Drug perturbation pilot study. [dataset] (2018), https://www.allencell.org/drug-perturbation-pilot.html, accessed 2026-02- 24

    2018

  2. [2]

    Nature Communi- cations15, 6516 (2024).https://doi.org/10.1038/s41467-024-50613-5

    Arevalo, J., Su, E., Ewald, J.D., van Dijk, R., Carpenter, A.E., Singh, S.: Evalu- ating batch correction methods for image-based cell profiling. Nature Communi- cations15, 6516 (2024).https://doi.org/10.1038/s41467-024-50613-5

    work page doi:10.1038/s41467-024-50613-5 2024

  3. [3]

    Biological Imaging4, e12 (Nov 2024).https://doi.org/10.1017/S2633903X2400014X

    Bendidi, I., Bardes, A., Cohen, E., Lamiable, A., Bollot, G., Genovesio, A.: Explor- ing self-supervised learning biases for microscopy image representation. Biological Imaging4, e12 (Nov 2024).https://doi.org/10.1017/S2633903X2400014X

    work page doi:10.1017/s2633903x2400014x 2024

  4. [4]

    Bourriez, N., Bendidi, I., Cohen, E., Watkinson, G., Sanchez, M., Bollot, G., Gen- ovesio, A.: Chada-vit : Channel adaptive attention for joint representation learning of heterogeneous microscopy images (2024),https://arxiv.org/abs/2311.15264

    Pith/arXiv arXiv 2024

  5. [5]

    In: Pei, J., Tseng, V.S., Cao, L., Motoda, H., Xu, G

    Campello, R.J.G.B., Moulavi, D., Sander, J.: Density-based clustering based on hierarchical density estimates. In: Pei, J., Tseng, V.S., Cao, L., Motoda, H., Xu, G. (eds.) Advances in Knowledge Discovery and Data Mining. pp. 160–172. Springer Berlin Heidelberg, Berlin, Heidelberg (2013) 16 J. Riel and V. M. Singh et al

    2013

  6. [6]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsuper- vised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neu- ral Information Processing Systems. vol. 33, pp. 9912–9924. Curran Associates, Inc. (2020),https://proceedings.neur...

    2020

  7. [7]

    In: ICCV (2021)

    Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)

    2021

  8. [8]

    bioRxiv (2023).https://doi.org/10.1101/2023.03.23.534023,https: //www.biorxiv.org/content/early/2023/03/24/2023.03.23.534023

    Chandrasekaran, S.N., Ackerman, J., Alix, E., Ando, D.M., Arevalo, J., Ben- nion, M., Boisseau, N., Borowa, A., Boyd, J.D., Brino, L., et al.: Jump cell painting dataset: morphological impact of 136,000 chemical and genetic pertur- bations. bioRxiv (2023).https://doi.org/10.1101/2023.03.23.534023,https: //www.biorxiv.org/content/early/2023/03/24/2023.03.23.534023

    doi:10.1101/2023.03.23.534023 2023

  9. [9]

    In: ICML (2020)

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: ICML (2020)

    2020

  10. [10]

    Nature Methods19(11), 1438–1448 (Nov 2022).https://doi.org/10.1038/s41592-022-01639-4,https: //doi.org/10.1038/s41592-022-01639-4

    Cutler, K.J., Stringer, C., Lo, T.W., Rappez, L., Stroustrup, N., Brook Peter- son, S., Wiggins, P.A., Mougous, J.D.: Omnipose: a high-precision morphology- independent solution for bacterial cell segmentation. Nature Methods19(11), 1438–1448 (Nov 2022).https://doi.org/10.1038/s41592-022-01639-4,https: //doi.org/10.1038/s41592-022-01639-4

    doi:10.1038/s41592-022-01639-4 2022

  11. [11]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248–255 (2009).https://doi.org/10.1109/CVPR.2009. 5206848

    doi:10.1109/cvpr.2009 2009

  12. [12]

    Tis- sue and Cell73, 101653 (2021).https://doi.org/https://doi.org/10.1016/ j.tice.2021.101653,https://www.sciencedirect.com/science/article/pii/ S0040816621001695

    Depto, D.S., Rahman, S., Hosen, M.M., Akter, M.S., Reme, T.R., Rah- man, A., Zunair, H., Rahman, M.S., Mahdy, M.: Automatic segmenta- tion of blood cells from microscopic slides: A comparative analysis. Tis- sue and Cell73, 101653 (2021).https://doi.org/https://doi.org/10.1016/ j.tice.2021.101653,https://www.sciencedirect.com/science/article/pii/ S0040816...

    arXiv 2021

  13. [13]

    Nature Communications11(1), 5723 (Nov 2020).https://doi.org/10.1038/ s41467-020-19557-4,https://doi.org/10.1038/s41467-020-19557-4

    Dietler, N., Minder, M., Gligorovski, V., Economou, A.M., Joly, D.A.H.L., Sadeghi, A., Chan, C.H.M., Kozi´ nski, M., Weigert, M., Bitbol, A.F., Rahi, S.J.: A convo- lutional neural network segments yeast microscopy images with high accuracy. Nature Communications11(1), 5723 (Nov 2020).https://doi.org/10.1038/ s41467-020-19557-4,https://doi.org/10.1038/s41...

    work page doi:10.1038/s41467-020-19557-4 2020

  14. [14]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  15. [15]

    In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining

    Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discov- ering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. p. 226–231. KDD’96, AAAI Press (1996)

    1996

  16. [16]

    In: European Congress on Digital Pathology

    Gamper, J., Koohbanani, N.A., Benes, K., Khuram, A., Rajpoot, N.: Pannuke: an open pan-cancer histology dataset for nuclei instance segmentation and classifica- tion. In: European Congress on Digital Pathology. pp. 11–19. Springer (2019)

    2019

  17. [17]

    arXiv preprint arXiv:2003.10778 (2020)

    Gamper, J., Koohbanani, N.A., Graham, S., Jahanifar, M., Khurram, S.A., Azam, A., Hewitt, K., Rajpoot, N.: Pannuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778 (2020)

    Pith/arXiv arXiv 2003

  18. [18]

    Goodman, A., Carpenter, A., Park, E., jlefman nvidia, BoozAllen, J., Kyle, Maggie, Nilofer, Sedivec, P., Cukierski, W.: 2018 data science bowl (2018), kaggle HASSL 17

    2018

  19. [19]

    Graham, S., Jahanifar, M., Vu, Q.D., Hadjigeorghiou, G., Leech, T., Snead, D., Raza, S.E.A., Minhas, F., Rajpoot, N.: Conic: Colon nuclei identification and counting challenge 2022 (2021),https://arxiv.org/abs/2111.14485

    Pith/arXiv arXiv 2022

  20. [20]

    In: Proceedings of the 34th International Conference on Neural Information Processing Systems

    Grill, J.B., Strub, F., Altch´ e, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Do- ersch, C., Pires, B.A., Guo, Z.D., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent a new approach to self-supervised learn- ing. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’...

    2020

  21. [21]

    Guo, Y., Xu, M., Li, J., Ni, B., Zhu, X., Sun, Z., Xu, Y.: Hcsc: Hierarchical contrastive selective coding. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022),https://openaccess.thecvf.com/ content/CVPR2022/papers/Guo_HCSC_Hierarchical_Contrastive_Selective_ Coding_CVPR_2022_paper.pdf

    2022

  22. [22]

    bioRxiv pp

    Gupta, A., Wefers, Z., Kahnert, K., Hansen, J.N., Leineweber, W., Cesnik, A., Lu, D., Axelsson, U., Ballllosera Navarro, F., Karaletsos, T., et al.: Subcell: Vision foundation models for microscopy capture single-cell biology. bioRxiv pp. 2024–12 (2024)

    2024

  23. [23]

    In: Medical Imaging with Deep Learning (MIDL), Proceedings of Machine Learning Research

    Haslum, J.F., Matsoukas, C., Leuchowius, K.J., M¨ ullers, E., Smith, K.: Metadata- guided consistency learning for high content images. In: Medical Imaging with Deep Learning (MIDL), Proceedings of Machine Learning Research. vol. 227, pp. 918–936 (2024),https://proceedings.mlr.press/v227/haslum24a.html

    2024

  24. [24]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2020),https: //arxiv.org/abs/1911.05722

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsu- pervised visual representation learning. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2020),https: //arxiv.org/abs/1911.05722

    Pith/arXiv arXiv 2020

  25. [25]

    Howard, A., Chow, A., CorporateResearchSartorius, Ca, M., Culliton, P., Jackson, T.: Sartorius - cell instance segmentation (2021), kaggle

    2021

  26. [26]

    npj Digital Medicine6(1), 74 (Apr 2023).https:// doi.org/10.1038/s41746-023-00811-0

    Huang, S.C., Pareek, A., Jensen, M., Lungren, M.P., Yeung, S., Chaudhari, A.S.: Self-supervised learning for medical image classification: a systematic review and implementation guidelines. npj Digital Medicine6(1), 74 (Apr 2023).https:// doi.org/10.1038/s41746-023-00811-0

    doi:10.1038/s41746-023-00811-0 2023

  27. [27]

    Hussein, N., Reinhard, B., Adrian, S., Marie-Lisa, E., Philipp, L., Katarzyna, B.: Lynsec: Lymphoma nuclear segmentation and classification (Jun 2023).https:// doi.org/10.5281/zenodo.8065174,https://doi.org/10.5281/zenodo.8065174

    work page doi:10.5281/zenodo.8065174 2023

  28. [28]

    In: NeurIPS (2020)

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS (2020)

    2020

  29. [29]

    Scientific Reports15(1), 4876 (2025)

    Kim, V., Adaloglou, N., Osterland, M., Morelli, F.M., Halawa, M., K¨ onig, T., Gnutt, D., Marin Zapata, P.A.: Self-supervision advances morphological profiling by unlocking powerful image representations. Scientific Reports15(1), 4876 (2025)

    2025

  30. [30]

    Kingma, D.P., Welling, M., et al.: Auto-encoding variational bayes (2013)

    2013

  31. [31]

    Nature Methods19(8), 995–1003 (Aug 2022).https://doi.org/10.1038/ s41592-022-01541-z

    Kobayashi, H., Cheveralls, K.C., Leonetti, M.D., Royer, L.A.: Self-supervised deep learning encodes high-resolution features of protein subcellular localiza- tion. Nature Methods19(8), 995–1003 (Aug 2022).https://doi.org/10.1038/ s41592-022-01541-z

    2022

  32. [32]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kraus, O., Kenyon-Dean, K., Saberian, S., Fallah, M., McLean, P., Leung, J., Sharma, V., Khan, A., Balakrishnan, J., Celik, S., et al.: Masked autoencoders for microscopy are scalable learners of cellular biology. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11757– 11768 (2024) 18 J. Riel and V. M. Singh et al

    2024

  33. [33]

    IEEE Transactions on Medical Imaging39(5), 1380–1391 (2020).https: //doi.org/10.1109/TMI.2019.2947628

    Kumar, N., Verma, R., Anand, D., et al.: A multi-organ nucleus segmentation challenge. IEEE Transactions on Medical Imaging39(5), 1380–1391 (2020).https: //doi.org/10.1109/TMI.2019.2947628

    doi:10.1109/tmi.2019.2947628 2020

  34. [34]

    In: 9th International Conference on Learning Represen- tations (ICLR) (2021)

    Li, J., Zhou, P., Xiong, C., Hoi, S.C.H.: Prototypical contrastive learning of unsu- pervised representations. In: 9th International Conference on Learning Represen- tations (ICLR) (2021)

    2021

  35. [35]

    Nature Methods21, 1103–1113 (2024).https://doi.org/https://doi.org/10

    Ma, J., Xie, R., Ayyadhury, S., Ge, C., Gupta, A., Gupta, R., Gu, S., Zhang, Y., Lee, G., Kim, J., Lou, W., Li, H., Upschulte, E., Dickscheid, T., de Almeida, J.G., Wang, Y., Han, L., Yang, X., Labagnara, M., Gligorovski, V., Scheder, M., Rahi, S.J., Kempster, C., Pollitt, A., Espinosa, L., Mignot, T., Middeke, J.M., Eckardt, J.N., Li, W., Li, Z., Cai, X....

    2024

  36. [36]

    Journal of machine learn- ing research9(Nov), 2579–2605 (2008)

    Maaten, L.v.d., Hinton, G.: Visualizing data using t-sne. Journal of machine learn- ing research9(Nov), 2579–2605 (2008)

    2008

  37. [37]

    arXiv preprint arXiv:2308.01760 (2023)

    Mahbod, A., Polak, C., Feldmann, K., Khan, R., Gelles, K., Dorffner, G., Woitek, R., Hatamikia, S., Ellinger, I.: Nuinsseg: A fully annotated dataset for nu- clei instance segmentation in h&e-stained histological images. arXiv preprint arXiv:2308.01760 (2023)

    Pith/arXiv arXiv 2023

  38. [38]

    Computers in Biology and Medicine132, 104349 (2021).https://doi.org/https://doi.org/10.1016/j

    Mahbod, A., Schaefer, G., Bancher, B., L¨ ow, C., Dorffner, G., Ecker, R., Ellinger, I.: Cryonuseg: A dataset for nuclei instance segmentation of cryosectioned h&e-stained histological images. Computers in Biology and Medicine132, 104349 (2021).https://doi.org/https://doi.org/10.1016/j. compbiomed.2021.104349,https://www.sciencedirect.com/science/article/...

    doi:10.1016/j 2021

  39. [39]

    IEEE Transactions on Med- ical Imaging38(2), 448–459 (Feb 2019).https://doi.org/10.1109/TMI.2018

    Naylor, P., La´ e, M., Reyal, F., Walter, T.: Segmentation of nuclei in histopathol- ogy images by deep regression of the distance map. IEEE Transactions on Med- ical Imaging38(2), 448–459 (Feb 2019).https://doi.org/10.1109/TMI.2018. 2865709,https://doi.org/10.1109/TMI.2018.2865709

    doi:10.1109/tmi.2018 2019

  40. [40]

    arXiv preprint arXiv:1807.03748 (2018)

    van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    Pith/arXiv arXiv 2018

  41. [41]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    Pith/arXiv arXiv 2023

  42. [42]

    bioRxiv (2025).https://doi.org/10

    Pachitariu, M., Rariden, M., Stringer, C.: Cellpose-sam: superhuman gen- eralization for cellular segmentation. bioRxiv (2025).https://doi.org/10. 1101/2025.04.28.651001,https://www.biorxiv.org/content/early/2025/05/ 01/2025.04.28.651001

    2025

  43. [43]

    Pfaendler, R.: Morphologically annotated single-cell images of human in- duced pluripotent stem cells for deep learning (Nov 2022).https://doi. org/10.3929/ethz-b-000581447,https://www.research-collection.ethz.ch/ entities/researchdata/b48aa89f-0c90-44e9-ad67-325ae37e3e89, data Collec- tion; Creative Commons Attribution-ShareAlike 4.0 International

    work page doi:10.3929/ethz-b-000581447 2022

  44. [44]

    Biorxiv pp

    Pfaendler, R., Hanimann, J., Lee, S., Snijder, B.: Self-supervised vision transform- ers accurately decode cellular state heterogeneity. Biorxiv pp. 2023–01 (2023)

    2023

  45. [45]

    Recursion Pharmaceuticals: OpenPhenom: Groundbreaking publicly accessible foundation models for microscopy data (2025), accessed October 31, 2025 HASSL 19

    2025

  46. [46]

    In: CVPR (2015).https://doi.org/10.1109/CVPR

    Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: CVPR (2015).https://doi.org/10.1109/CVPR. 2015.7298682

    doi:10.1109/cvpr 2015

  47. [47]

    ethz.ch/entities/researchdata/d0feef10-453f-4b8d-b8e1-582ea2976f91, dataset; Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

    Severin, Y.: Deep phenotyping reveals the molecular and health de- terminants of human immune cell morphology (Jun 2021).https: //doi.org/10.3929/ethz-b-000343106,https://www.research-collection. ethz.ch/entities/researchdata/d0feef10-453f-4b8d-b8e1-582ea2976f91, dataset; Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

    work page doi:10.3929/ethz-b-000343106 2021

  48. [48]

    arXiv preprint arXiv:2508.10104 (2025).https://doi.org/10.48550/ arXiv.2508.10104

    Sim´ eoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., J´ egou, H., Labatut, P., Bojanowski, P.: DINOv3. arXiv prepri...

    Pith/arXiv arXiv 2025

  49. [49]

    bioRxiv (2020).https://doi.org/10

    Stringer, C., Michaelos, M., Pachitariu, M.: Cellpose: a generalist al- gorithm for cellular segmentation. bioRxiv (2020).https://doi.org/10. 1101/2020.02.02.931238,https://www.biorxiv.org/content/early/2020/02/ 03/2020.02.02.931238

    2020

  50. [50]

    Advances in neural information processing systems (2017)

    Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems (2017)

    2017

  51. [51]

    Accessed 2025-10-27

    Van Valen Lab (Caltech): Tissuenet: training dataset for nuclear and whole-cell segmentation (Apr 2022), deepCell datasets documentation; modified Apache li- cense (non-commercial academic use). Accessed 2025-10-27

    2022

  52. [52]

    html, deepCell Datasets; non-commercial academic use (modified Apache license); accessed 2025-10-27

    Van Valen Lab (Caltech): Dynamicnuclearnet segmentation (v1.0) (2023),https: //deepcell.readthedocs.io/en/master/data-gallery/dynamicnuclearnet. html, deepCell Datasets; non-commercial academic use (modified Apache license); accessed 2025-10-27

    2023

  53. [53]

    IEEE Transactions on Medical Imaging40(12), 3413–3423 (2021).https://doi.org/10.1109/TMI.2021.3085712

    Verma, R., Kumar, N., et al.: Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transactions on Medical Imaging40(12), 3413–3423 (2021).https://doi.org/10.1109/TMI.2021.3085712

    doi:10.1109/tmi.2021.3085712 2021

  54. [54]

    Vinh, N.X., Epps, J., Bailey, J.: Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance. J. Mach. Learn. Res.11, 2837–2854 (Dec 2010)

    2010

  55. [55]

    Frontiers in Bioengineering and Biotech- nology7, 53 (Apr 2019).https://doi.org/10.3389/fbioe.2019.00053,https: //www.frontiersin.org/articles/10.3389/fbioe.2019.00053/full

    Vu, Q.D., Graham, S., Kurc, T., To, M.N.N., Shaban, M., Qaiser, T., Koohbanani, N.A., Khurram, S.A., Kalpathy-Cramer, J., Zhao, T., Gupta, R., Kwak, J.T., Rajpoot, N., Saltz, J., Farahani, K.: Methods for segmentation and classifica- tion of digital microscopy tissue images. Frontiers in Bioengineering and Biotech- nology7, 53 (Apr 2019).https://doi.org/1...

    doi:10.3389/fbioe.2019.00053 2019

  56. [56]

    Biomedical Signal Processing and Control93, 106143 (2024)

    Wang, R., Qiu, Y., Hao, X., Jin, S., Gao, J., Qi, H., Xu, Q., Zhang, Y., Xu, H.: Simultaneously segmenting and classifying cell nuclei by using multi-task learning in multiplex immunohistochemical tissue microarray sections. Biomedical Signal Processing and Control93, 106143 (2024)

    2024

  57. [57]

    Winsnes, C., Lundberg, E., Maggie, Culliton, P., Le, T., UAxelsson, Ouyang, W.: Human protein atlas - single cell classification (2021), kaggle

    2021

  58. [58]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018),https://arxiv

    Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via non- parametric instance discrimination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018),https://arxiv. org/abs/1805.01978 20 J. Riel and V. M. Singh et al

    Pith/arXiv arXiv 2018

  59. [59]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(3), 3753– 3767 (2023).https://doi.org/10.1109/TPAMI.2022.3176690,https://arxiv

    Xu, H., Zhang, X., Li, H., Xie, L., Dai, W., Xiong, H., Tian, Q.: Seed the views: Hierarchical semantic alignment for contrastive representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence45(3), 3753– 3767 (2023).https://doi.org/10.1109/TPAMI.2022.3176690,https://arxiv. org/abs/2012.02733

    work page doi:10.1109/tpami.2022.3176690 2023

  60. [60]

    Scientific Data10(1), 41 (Jan 2023).https://doi.org/10.1038/ s41597-022-01721-8,https://doi.org/10.1038/s41597-022-01721-8

    Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., Ni, B.: Medm- nist v2 – a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data10(1), 41 (Jan 2023).https://doi.org/10.1038/ s41597-022-01721-8,https://doi.org/10.1038/s41597-022-01721-8

    doi:10.1038/s41597-022-01721-8 2023

  61. [61]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yao, H., Hanslovsky, P., Huetter, J.C., Hoeckendorf, B., Richmond, D.: Weakly supervised set-consistency learning improves morphological profiling of single-cell images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6978–6987 (2024)

    2024

  62. [62]

    Znale´ zniak, M., Rola, P., Kaszuba, P., Tabor, J.,´Smieja, M.: Contrastive hierarchi- cal clustering. arXiv preprint arXiv:2303.03389 (2023),https://arxiv.org/abs/ 2303.03389 HASSL 1 HASSL: Hierarchy-Aware Self-Supervised Learning Framework for Single Cell Microscopy Supplementary Material A Loss Calculation Algorithms To aid in the understanding of the ...

    Pith/arXiv arXiv 2023