Pith. sign in

REVIEW 5 major objections 5 minor 43 references

ST-DAI: Single-shot 2.5D Spatial Transcriptomics with Intra-Sample Domain Adaptive Imputation for Cost-efficient 3D Reconstruction

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ST-DAI reconstructs 3D spatial transcriptomics from one fully sequenced section and sparse samples of adjacent sections, matching fully sampled fidelity while cutting sequencing cost.

desk verdict New 2.5D sampling plus intra-sample adaptation idea, but the headline cost-saving claim rests on a metric artifact (DCO) and an untested fully-sampled oracle; worth reviewing, not yet convincing. read the letter →

arxiv 2507.21516 v1 pith:O2WOL6OM submitted 2025-07-29 eess.IV cs.CV

Jiahe Qian , Yaoyu Fang , Xinkun Wang , Lee A. Cooper , Bo Zhou This is my paper · ORCID
classification eess.IVcs.CV
keywords spatialtranscriptomics3Dreconstructiongeneexpressionimputationdomainadaptationpseudo-labelingparameter-efficientfine-tuning2.5Dsampling
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

This paper claims that a full 3D spatial-transcriptomics volume can be reconstructed from one fully sequenced tissue section plus sparse measurements on its neighbors, using only data from the sample itself. The authors argue that this removes the need for large external training corpora, which are costly to assemble and generalize poorly across tissue domains. They propose a sampling scheme that sequences a central section completely and measures a quarter of the spots on adjacent sections, then an imputation pipeline that aligns sections, generates confidence-weighted pseudo-labels, and adapts a pretrained network to each adjacent section with a small number of trainable parameters. On six breast-cancer section pairs, the reconstructed gene-expression maps match or exceed the fidelity of fully sampled baselines as measured by PSNR, SSIM, MAE, and correlation. If the result holds, 3D transcriptomic atlases could be produced at roughly a quarter of the sequencing cost as the number of sections grows.

What carries the argument

The load-bearing mechanism is a four-stage pipeline: Cross-section Alignment (CSA) uses SIFT keypoints and RANSAC to register the central section to each adjacent section; a Pseudo Map Network pretrained on the central section's masked expression generates dense pseudo-labels for adjacent sections; Fast Multi-Domain Refinement (FMDR) fine-tunes only Parameter-Efficient Domain-Alignment Layers (PDLs, channel-wise affine transforms initialized to identity, $\hat{f} = (1+a)\odot f + b$) plus the terminal layer while freezing the shared backbone; and a Confidence Score Generator (CSG) reweights pseudo-labels by interpolating errors at the sparse observed locations, with a Data Consistency Operation (DCO) replacing predicted values at sampled spots with measured values at inference. The identity-initialized PDL keeps the pretrained feature distribution intact while allowing cheap per-section calibration.

What would settle it

Fully sequence an adjacent section as ground truth, run ST-DAI using only the 2x2 sparse grid, and compare the imputed values with the held-out full map; the central claim fails if the imputation error is no better than a model trained only on the central section, or if corrupting the pseudo-labels by pretraining on a deliberately mismatched section does not degrade the final reconstruction.

Watch

Extended reading notes

Core claim

The central claim is that intra-sample domain adaptation, not external training data, is sufficient to turn a 2.5D acquisition into a dense 3D expression map. After registering the fully sampled central section to a sparsely sampled neighbor, a network pretrained on the central section generates pseudo-labels for the neighbor's unmeasured spots, and refining only a few domain-alignment parameters with confidence-weighted pseudo-labels recovers expression patterns that match fully sampled ground truth. The paper reports consistent gains across all tested section pairs and with two different backbone architectures, and attributes the gains specifically to the alignment, confidence reweighting, and parameter-efficient refinement modules rather than to the backbone architecture itself.

Load-bearing premise

The whole refinement step rests on trusting the computer-generated stand-in values for the unmeasured spots on adjacent sections, and those stand-ins come from a network trained on the central section; if that network is systematically wrong, the final reconstruction will be systematically wrong too.

Editorial extensions

If this is right

  • 3D spatial transcriptomics can be produced without sequencing every serial section, cutting per-sample sequencing cost to roughly 25% as the number of sections grows.
  • No external image-transcriptome training corpus is needed; the model is calibrated on the sample itself, avoiding cross-sample domain shift.
  • The framework works with both UNet and diffusion backbones, so it is architecture-agnostic.
  • Confidence-weighted pseudo-labeling plus parameter-efficient refinement recovers fine-grained spatial patterns that generic imputation baselines blur.
  • Each adjacent section is handled independently, so the method extends to stacks of arbitrary depth at linear cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the uniform 2x2 grid is likely not the optimal sampling allocation; an adaptive sampler that spends more measurements in regions where pseudo-label confidence is low could push the cost below 25% or improve fidelity at the same cost.
  • Beyond the paper: the same alignment-pseudo-label-confidence-refinement loop should transfer to other serial-section imaging modalities, such as protein or metabolite imaging, where adjacent sections share morphology but differ in domain.
  • Beyond the paper: the pseudo-label bottleneck suggests a direct test of how reconstruction quality degrades as the central-to-adjacent tissue distance, and hence domain shift, increases; the method's practical ceiling is set by how far a single reference section can be extrapolated.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes ST-DAI, a framework for reconstructing three-dimensional spatial transcriptomics from a '2.5D' sampling scheme in which one central tissue section is fully sequenced and adjacent sections are sparsely sequenced on a regular 25% grid. The method combines cross-section alignment (CSA), a pseudo-map network pretrained on the central section, a fast multi-domain refinement stage with parameter-efficient domain-alignment layers (PDLs), a confidence score generator (CSG), and a data consistency operation (DCO) that overwrites predictions at measured pixels with the observed values. Experiments on six Xenium breast-cancer sections arranged into three adjacent pairs report consistent improvements in PSNR, SSIM, MAE, and PCC over vanilla UNet/Diffusion backbones and over TESLA and DIST baselines, with ablations attributed to each proposed module.

Significance. If the central claim were established, ST-DAI would offer a practical way to reduce sequencing cost for 3D spatial transcriptomics to roughly 25% of full sampling while retaining accurate reconstruction, and its sample-specific training would avoid the cross-sample domain-shift problems of external-dataset methods. The parameter-efficient PDL design is a reasonable practical choice, and the ablation study is informative in showing that each learned component contributes to the reported numbers. However, the current evaluation does not substantiate the headline claims: the reported margins are confounded by the data consistency operation, no fully-sampled oracle is tested, and the pseudo-label supervision is a self-distillation loop whose quality is never directly validated. The underlying idea is promising, but the evidence as presented is not yet sufficient to support the cost-saving and parity claims.

major comments (5)
  1. [§2.5, Eq. (3); Table 3] The data consistency operation confounds the evaluation. DCO replaces every prediction at the 25% sampled pixels with the exact measured value, and because the reported metrics are computed over all pixels, roughly one quarter of the evaluated pixels are perfect by construction. In Table 3, moving from row 6 (CSA+FMDR+PDL+CSG, no DCO) to row 7 (full ST-DAI) increases PSNR by 2.45 dB on the TENX95/TENX97 pair, whereas the cumulative gain of all learned components over the vanilla baseline is only 2.15 dB. The baselines in Tables 1 and 2 are not described as receiving DCO, so their metrics include errors at observed pixels that ST-DAI eliminates by construction. The comparison is therefore not apples-to-apples, and the large reported margins may largely reflect the evaluation protocol rather than the proposed imputation modules. Please report metrics on the unsampled pixels only, apply DCO to all compared methods, or both.
  2. [Abstract; §3.1, Tables 1–2] The abstract claims that ST-DAI achieves performance 'comparable to fully sampled approaches,' but no fully-sampled oracle is tested. The experiments compare against backbones trained on sparse data and against TESLA and DIST, none of which is a fully-sampled baseline. To support the central cost-saving claim, the authors should include an oracle baseline such as the same UNet trained with full ground-truth supervision on the adjacent section, and report metrics on unsampled pixels for that oracle as well. Without such a comparison, the phrase 'comparable to fully sampled approaches' is unsupported.
  3. [§2.2, Eq. (6); §2.3] The refinement stage trains the adjacent-section branch to match pseudo-labels produced by a network pretrained on the central section, and the confidence weighting in CSG uses ground truth only at the sparsely sampled observed pixels. This is a form of self-distillation: if the pretrained network has systematic biases, the refinement can reinforce those biases rather than correct them. The manuscript does not directly evaluate pseudo-label quality, for example by holding out observed pixels during pretraining and measuring pseudo-label error on them, or by cross-validating the final imputation on held-out observed sites. Please add such an evaluation, and additionally report the imputation metrics on unsampled pixels only, where pseudo-label quality is actually tested.
  4. [§2.8, Tables 1–2] The baseline setup for TESLA and DIST is under-specified. It is unclear whether these methods receive the same sparse 25% measurements, the same histology images, the same cross-section alignment, or the same DCO post-processing. TESLA and DIST were designed for different tasks (sub-spot super-resolution and denoising, respectively), so a fair comparison requires a clear statement of how they were adapted to the 2.5D sampling setting, including any hyperparameters and input features. Without a controlled protocol, the large performance gaps in Table 2 cannot be attributed to the proposed method rather than to differences in the information each baseline was allowed to use.
  5. [§3.1, Table 1] The evaluation covers only three section pairs and eight genes, and the reported means are accompanied by large standard deviations across genes (e.g., PSNR 27.21 ± 3.45 dB for TENX95/TENX97). No statistical significance testing or per-gene breakdown is provided, so the claim of 'consistent' improvement across diverse samples is not strongly supported. At minimum, please report per-gene results or a paired statistical test, and discuss the range of gains across genes and section pairs.
minor comments (5)
  1. [§2.8] Reference [22] points to the 10x Genomics Xenium web page, but the text says the data come from the HEST–1K dataset; the HEST–1K paper should be cited instead.
  2. [§2.4] There is a typo, 'subtal and precise domain alignment,' which should read 'subtle and precise.'
  3. [§2.2, §5] There are grammatical issues such as 'a intra-sample domain-adaptive imputation framework' and 'a intra-sample domain-adaptive imputation pipeline'; 'a' should be 'an.'
  4. [§2.8] Please clarify whether all metrics are computed only within the tissue mask or over the entire image, and specify the normalization applied to gene-expression values before computing PSNR and SSIM.
  5. [Table 1] The table caption states 'mean ± SD' without defining that the standard deviation is across genes; please state this explicitly in the caption.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported gains are partly forced by construction: DCO injects the ground-truth values at the 25% sampled pixels into the final output, and the aggregate metrics reward those pixels exactly.

  1. self definitional [Section 2.5, Eq. (3); Table 3 rows 6-7]
    "For each pixel i in an adjacent section a, let G a i denote the measured expression value when pixel i is sampled, and let G a pred,i denote the network prediction. DCO is defined as G a final(i)= G a i , if pixel i is sampled, G a pred,i, otherwise. (3) This rule guarantees that the final output G a final faithfully preserves all real measurements, while filling unsampled regions with the network's predictions."

    The reported PSNR/SSIM/MAE/PCC in Tables 1-3 are computed over the entire adjacent section against its fully measured ground-truth map. At the 25% pixels on the 2x2 sampling grid, Eq. (3) sets G_a_final(i) = G_a_i, i.e., the output is the ground-truth value itself, so these pixels contribute zero error to every metric by construction. The ablation quantifies the effect: row 6 (all learned modules, no DCO) gives 24.76 dB PSNR; row 7 (adding DCO) gives 27.21 dB, a 2.45 dB jump that exceeds the cumulative 2.15 dB contributed by CSA+FMDR+PDL+CSG over the vanilla baseline. Baselines TESLA, DIST, and vanilla UNet receive no DCO, so part of the large margin in Tables 1-2 is protocol leakage rather than learned imputation.

full rationale

No load-bearing self-citation chain or imported uniqueness theorem exists. The pseudo-label supervision in Eq. (6) is a self-training design, but it is not circular for the final claim because the reported evaluation uses real, fully measured adjacent-section expression as ground truth, not the pseudo-labels. However, the headline 'prediction performance comparable to fully sampled approaches' is partly forced by construction: the Data Consistency Operation (Eq. 3) copies experimentally observed values into the final output at sampled pixels, and the aggregate metrics in Tables 1-3 average over those pixels. Table 3 shows DCO is the largest single contributor (2.45 dB PSNR) and is not applied to the TESLA, DIST, or vanilla UNet baselines, so the comparison does not isolate imputation accuracy on the unsampled majority of the tissue. A clean evaluation would report metrics only on unsampled pixels or would apply DCO to all methods. Since the core learned imputation is still tested against real data and the circularity is confined to the evaluation protocol, the overall score is 6 rather than higher.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that a central section can serve as a reliable teacher for its neighbors, and on several standard computer vision and deep learning premises. No new physical entities or constants are introduced.

free parameters (3)
  • Sparse sampling grid spacing = 2 pixels in both axial directions
    Chosen in Section 2.1 to set 25% sampling density; no sweep or justification, and it controls the difficulty and cost of imputation.
  • Confidence map renormalization = unit mean
    Section 2.3 normalizes confidence weights to unit mean; this scaling is a design choice to balance loss terms, not derived from data.
  • Training duration = 500 to 2000 epochs depending on stage and backbone
    Section 2.7 fixes these values; no early stopping criteria or sensitivity analysis.
assumptions (4)
  • domain assumption Adjacent sections from the same tissue share enough morphological and transcriptomic similarity that a network trained on the central section produces useful pseudo-labels for adjacent sections.
    Stated in Section 1 (intra-sample domain gap) and used throughout; only tested on 3 breast cancer pairs.
  • standard math SIFT keypoints with RANSAC give a transformation that aligns consecutive tissue sections adequately.
    Used in CSA (Section 2.2); standard computer vision pipeline but can fail under large deformation or low texture.
  • domain assumption Gene expression imputation is a well-posed regression problem from histology plus local sparse expression.
    Core premise of the imputation network; relies on known but imperfect histology-expression correlation.
  • standard math L2 loss is an appropriate objective for gene expression imputation.
    Used in Eqs. (4) to (6); expression counts are non-negative and skewed, so L2 may not be optimal but is conventional.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ST-DAI: Single-shot 2.5D Spatial Transcriptomics with Intra-Sample Domain Adaptive Imputation for Cost-efficient 3D Reconstruction." pith.science (2026). https://pith.science/paper/O2WOL6OM

@misc{pith2026250721516,
  author       = {Pith},
  title        = {Pith review of: ST-DAI: Single-shot 2.5D Spatial Transcriptomics with Intra-Sample Domain Adaptive Imputation for Cost-efficient 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O2WOL6OM}},
  note         = {Machine review of arXiv:2507.21516}
}
read the original abstract

For 3D spatial transcriptomics (ST), the high per-section acquisition cost of fully sampling every tissue section remains a significant challenge. Although recent approaches predict gene expression from histology images, these methods require large external datasets, which leads to high-cost and suffers from substantial domain discrepancies that lead to poor generalization on new samples. In this work, we introduce ST-DAI, a single-shot framework for 3D ST that couples a cost-efficient 2.5D sampling scheme with an intra-sample domain-adaptive imputation framework. First, in the cost-efficient 2.5D sampling stage, one reference section (central section) is fully sampled while other sections (adjacent sections) is sparsely sampled, thereby capturing volumetric context at significantly reduced experimental cost. Second, we propose a single-shot 3D imputation learning method that allows us to generate fully sampled 3D ST from this cost-efficient 2.5D ST scheme, using only sample-specific training. We observe position misalignment and domain discrepancy between sections. To address those issues, we adopt a pipeline that first aligns the central section to the adjacent section, thereafter generates dense pseudo-supervision on the central section, and then performs Fast Multi-Domain Refinement (FMDR), which adapts the network to the domain of the adjacent section while fine-tuning only a few parameters through the use of Parameter-Efficient Domain-Alignment Layers (PDLs). During this refinement, a Confidence Score Generator (CSG) reweights the pseudo-labels according to their estimated reliability, thereby directing imputation toward trustworthy regions. Our experimental results demonstrate that ST-DAI achieves gene expression prediction performance comparable to fully sampled approaches while substantially reducing the measurement burden.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [1]

    Nature methods 18(1), 9–14 (2021)

    Marx, V.: Method of the year: spatially resolved transcriptomics. Nature methods 18(1), 9–14 (2021)

    work page 2021

  2. [2]

    Science 353(6294), 78–82 (2016)

    St ˚ ahl, P.L., Salm´ en, F., Vickovic, S., Lundmark, A., Navarro, J.F., Magnusson, J., Giacomello, S., Asp, M., Westholm, J.O., Huss, M., et al.: Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science 353(6294), 78–82 (2016)

    work page 2016

  3. [3]

    Science 363(6434), 1463–1467 (2019)

    Rodriques, S.G., Stickels, R.R., Goeva, A., Martin, C.A., Murray, E., Vanderburg, C.R., Welch, J., Chen, L.M., Chen, F., Macosko, E.Z.: Slide-seq: A scalable tech- nology for measuring genome-wide expression at high spatial resolution. Science 363(6434), 1463–1467 (2019)

    work page 2019

  4. [4]

    Science 361(6400), 5691 (2018)

    Wang, X., Allen, W.E., Wright, M.A., Sylwestrak, E.L., Samusik, N., Vesuna, S., Evans, K., Liu, C., Ramakrishnan, C., Liu, J., et al.: Three-dimensional intact- tissue sequencing of single-cell transcriptional states. Science 361(6400), 5691 (2018)

    work page 2018

  5. [5]

    Nature communications 11(1), 3877 (2020)

    Schmauch, B., Romagnoni, A., Pronier, E., Saillard, C., Maill´ e, P., Calderaro, J., Kamoun, A., Sefta, M., Toldo, S., Zaslavskiy, M., et al.: A deep learning model to predict rna-seq expression of tumours from whole slide images. Nature communications 11(1), 3877 (2020)

    work page 2020

  6. [6]

    Advances in Neural Information Processing Systems 36, 70626–70637 (2023)

    Xie, R., Pang, K., Chung, S., Perciani, C., MacParland, S., Wang, B., Bader, 17 G.: Spatially resolved gene expression prediction from histology images via bi- modal contrastive learning. Advances in Neural Information Processing Systems 36, 70626–70637 (2023)

    work page 2023

  7. [7]

    Nature medicine 24(10), 1559–1567 (2018)

    Coudray, N., Ocampo, P.S., Sakellaropoulos, T., Narula, N., Snuderl, M., Feny¨ o, D., Moreira, A.L., Razavian, N., Tsirigos, A.: Classification and mutation predic- tion from non–small cell lung cancer histopathology images using deep learning. Nature medicine 24(10), 1559–1567 (2018)

    work page 2018

  8. [8]

    Nature medicine 25(8), 1301–1309 (2019)

    Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical- grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine 25(8), 1301–1309 (2019)

    2019

Show all 43 references
  1. [9]

    arXiv preprint arXiv:2412.03026 (2024)

    Zhu, J., Deng, R., Yao, T., Xiong, J., Qu, C., Guo, J., Lu, S., Yin, M., Wang, Y., Zhao, S., Yang, H., Huo, Y.: Asign: An anatomy-aware spatial imputation graphic network for 3d spatial transcriptomics. arXiv preprint arXiv:2412.03026 (2024)

    2024 arXiv

  2. [10]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Yang, Y., Hossain, M.Z., Stone, E.A., Rahman, S.: Exemplar guided deep neural network for spatial transcriptomics analysis of gene expression prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 5039–5048 (2023)

    2023

  3. [11]

    Science Advances 6(26), 3446 (2020) https://doi.org/10.1126/sciadv.abb3446

    Ortiz, C., Navarro, J.F., Jurek, A., Martin, A., Lundeberg, J., Meletis, K.: Molecular atlas of the adult mouse brain. Science Advances 6(26), 3446 (2020) https://doi.org/10.1126/sciadv.abb3446

    2020 doi

  4. [12]

    Nature Communications 12(1), 6012 (2021) https://doi.org/10.1038/ s41467-021-26271-2

    Andersson, A., Larsson, L., Stenbeck, L., Salmen, F., Ehinger, A., Wu, S.Z., Al-Eryani, G., Roden, D., Swarbrick, A., Borg, A.e.a.: Spatial deconvolution of her2-positive breast cancer delineates tumor-associated cell type interac- tions. Nature Communications 12(1), 6012 (202...

    2021

  5. [13]

    Nature Genetics 53(9), 1334–1347 (2021) https://doi.org/10.1038/s41588-021-00904-2

    Wu, S.Z., Al-Eryani, G., Roden, D.L., Junankar, S., Harvey, K., Andersson, A., Thennavan, A., Wang, C., Torpy, J.R., Bartonicek, N.e.a.: A single-cell and spa- tially resolved atlas of human breast cancers. Nature Genetics 53(9), 1334–1347 (2021) https://doi.org/10.1038/s41588...

    2021 doi

  6. [14]

    Medical image analysis 58, 101544 (2019)

    Tellez, D., Litjens, G., B´ andi, P., Bulten, W., Bokhorst, J.-M., Ciompi, F., Van Der Laak, J.: Quantifying the effects of data augmentation and stain color nor- malization in convolutional neural networks for computational pathology. Medical image analysis 58, 101544 (2019)

    2019

  7. [15]

    Nature communications 14(1), 3459 18 (2023)

    Saillard, C., Delecourt, F., Schmauch, B., Moindrot, O., Svrcek, M., Bardier- Dupas, A., Emile, J.F., Ayadi, M., Rebours, V., Mestier, L., et al.: Pacpaint: a histology-based deep learning model uncovers the extensive intratumor molecular heterogeneity of pancreatic adenocarci...

    2023

  8. [16]

    IEEE journal of biomedical and health informatics 25(2), 325–336 (2020)

    Stacke, K., Eilertsen, G., Unger, J., Lundstr¨ om, C.: Measuring domain shift for deep learning in histopathology. IEEE journal of biomedical and health informatics 25(2), 325–336 (2020)

    2020

  9. [17]

    Interna- tional journal of computer vision 60, 91–110 (2004)

    Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Interna- tional journal of computer vision 60, 91–110 (2004)

    2004

  10. [18]

    FISCHLER AND, M.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381–395 (1981)

    1981

  11. [19]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention, pp

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 234–241 (2015). Springer

    2015

  12. [20]

    arXiv preprint arXiv:2006.11239 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arXiv:2006.11239 (2020)

    2020 arXiv

  13. [21]

    Nature Medicine 30(3), 863–874 (2024)

    Lu, M.Y., Chen, B., Williamson, D.F., Chen, R.J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Le, L.P., Gerber, G., et al.: A visual-language foundation model for computational pathology. Nature Medicine 30(3), 863–874 (2024)

    2024

  14. [22]

    Available at: https://www

    10x Genomics: Xenium: Spatial Gene Expression. Available at: https://www. 10xgenomics.com/products/xenium (2022)

    2022

  15. [23]

    Cell systems 14(5), 404–417 (2023)

    Hu, J., Coleman, K., Zhang, D., Lee, E.B., Kadara, H., Wang, L., Li, M.: Deci- phering tumor ecosystems at super resolution from spatial transcriptomics with tesla. Cell systems 14(5), 404–417 (2023)

    2023

  16. [24]

    Briefings in Bioinformatics 24(2), 013 (2023)

    Zhao, Y., Wang, K., Hu, G.: Dist: spatial transcriptomics enhancement using deep learning. Briefings in Bioinformatics 24(2), 013 (2023)

    2023

  17. [25]

    arXiv preprint arXiv:2402.12187 (2024)

    Park, L.H., Kim, J., Oh, M.G., Park, J., Kwon, T.: Adversarial feature alignment: Balancing robustness and accuracy in deep learning via adversarial training. arXiv preprint arXiv:2402.12187 (2024)

    2024 arXiv

  18. [26]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Li, Y., Zhang, Z., Yang, Y., Xu, D., Yang, M.-H.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5400–5409 (2018)

    2018

  19. [27]

    arXiv preprint arXiv:2109.09033 (2021)

    Zhang, B., Chen, T., Wang, B., Li, R.: Joint distribution alignment via adversarial learning for domain adaptive object detection. arXiv preprint arXiv:2109.09033 (2021)

    2021 arXiv

  20. [28]

    Pattern Recognition 108, 107539 (2020) 19

    Zhang, Z., Zhang, Z., Zhang, X., Wang, Z., Li, Z.: Adversarial feature distribution alignment for semi-supervised learning. Pattern Recognition 108, 107539 (2020) 19

    2020

  21. [29]

    In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pp

    Wang, H., Shen, T., Zhang, W., Duan, L., Mei, T.: An adversarial domain adap- tation network for cross-domain fine-grained recognition. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pp. 1244–1252 (2020)

    2020

  22. [30]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 39(8), 1634–1640 (2017)

    Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(8), 1634–1640 (2017)

    2017

  23. [31]

    In: Proceedings of the 36th Conference on Neural Information Processing Systems, pp

    Jang, J., Na, B., Shin, D., Ji, M., Song, K., Moon, I.-C.: Unknown-aware domain adversarial learning for open-set domain adaptation. In: Proceedings of the 36th Conference on Neural Information Processing Systems, pp. 1–13 (2022)

    2022

  24. [32]

    arXiv preprint arXiv:2503.21695 (2025)

    Qian, J., Fang, Y., Hao, J., Zhou, B.: Ama-sam: Adversarial multi-domain align- ment of segment anything model for high-fidelity histology nuclei segmentation. arXiv preprint arXiv:2503.21695 (2025)

    2025 arXiv

  25. [33]

    arXiv preprint arXiv:2502.00052 (2025)

    Quintana, G.I., Vancamberg, L., Jugnon, V., Desolneux, A., Mougeot, M.: Bridg- ing contrastive learning and domain adaptation: Theoretical perspective and practical application. arXiv preprint arXiv:2502.00052 (2025)

    2025 arXiv

  26. [34]

    arXiv preprint arXiv:2103.08454 (2021)

    Liu, Z., Zhu, Z., Zheng, S., Liu, Y., Zhou, J., Zhao, Y.: Margin preserving self-paced contrastive learning towards domain adaptation for medical image segmentation. arXiv preprint arXiv:2103.08454 (2021)

    2021 arXiv

  27. [35]

    In: Proceedings of the Fifth Conference on Health, Inference, and Learning, pp

    En, Q., Guo, Y.: Unsupervised domain adaptation for medical image segmenta- tion with dynamic prototype-based contrastive learning. In: Proceedings of the Fifth Conference on Health, Inference, and Learning, pp. 312–325. PMLR, ??? (2024)

    2024

  28. [36]

    arXiv preprint arXiv:2208.08605 (2022)

    Gu, R., Zhang, J., Wang, G., Lei, W., Song, T., Zhang, X., Li, K., Zhang, S.: Contrastive semi-supervised learning for domain adaptive segmentation across similar anatomical structures. arXiv preprint arXiv:2208.08605 (2022)

    2022 arXiv

  29. [37]

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

    Thota, M., Leontidis, G.: Contrastive domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 1110–1119 (2021)

    2021

  30. [38]

    Nature Machine Intelligence 5(8), 487–499 (2023)

    Tang, J., Zhang, W., Li, Y., Liu, Y., Zhang, Z.: Self-supervised learning for medical image classification. Nature Machine Intelligence 5(8), 487–499 (2023)

    2023

  31. [39]

    In: Proceedings of the 39th International Conference on Machine Learning, pp

    Shen, K., Jones, R., Kumar, A., Xie, S.M.: Explaining contrastive learning for unsupervised domain adaptation. In: Proceedings of the 39th International Conference on Machine Learning, pp. 10128–10146. PMLR, ??? (2022)

    2022

  32. [40]

    In: Proceedings of the 20 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Lee, H.J., Kim, J.U., Lee, S., Kim, H.G., Ro, Y.M.: Structure boundary preserving segmentation for medical image with ambiguous boundary. In: Proceedings of the 20 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1–9 (2020)

    2020

  33. [41]

    Journal of Medical Imaging 8(4), 1–12 (2021)

    Chen, W., Zhang, Y., Li, W., Liu, Z., Wang, J.: Medical image mis-segmentation region refinement framework based on dynamic graph convolution. Journal of Medical Imaging 8(4), 1–12 (2021)

    2021

  34. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    El-Jurdi, N., Sadiq, M., Al-Kadi, O., Al-Masni, M.A., Al-Maadeed, S.: A sur- prisingly effective perimeter-based loss for medical image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1–9 (2021)

    2021

  35. [43]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Hatamizadeh, A., Terzopoulos, D., Myronenko, A.: End-to-end boundary aware networks for medical image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1–9 (2019) 21

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.