Pith. sign in

REVIEW 4 major objections 4 minor 38 references

OTSurv: A Novel Multiple Instance Learning Framework for Survival Prediction with Heterogeneity-aware Optimal Transport

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read OTSurv claims that WSI survival prediction is best framed as a heterogeneity-aware optimal transport problem, and reports a new state-of-the-art mean C-index of 0.646 across six TCGA benchmarks.

desk verdict Solid OT-MIL engineering, but the headline SOTA claim is not supported by the paper's own Table 1. read the letter →

arxiv 2506.20741 v2 pith:RFVIBODZ submitted 2025-06-25 cs.CV

classification cs.CV
keywords survivalpredictionmultipleinstancelearningoptimaltransportwholeslideimagespathologicalheterogeneityC-indexunbalancedTCGAbenchmarks
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

OTSurv argues that survival prediction from whole-slide images is best modeled as an optimal transport problem: align a variable-sized bag of patch features to a fixed set of learned survival tokens, and let the transport plan itself define which patches matter. The paper adds two marginal constraints to that plan — a global penalty on the token mass distribution and a local, curriculum-style total-mass constraint — and rewrites the constrained problem as an unbalanced optimal transport problem with a virtual token, solvable by a differentiable matrix scaling algorithm. On six TCGA cancer benchmarks the method reports a new best average C-index, 0.646 versus 0.610 for the previous state of the art, with statistically significant log-rank splits. A sympathetic reader would take away that explicitly structuring the aggregation objective as transport with marginal constraints can preserve rare prognostic regions and suppress noisy patches better than attention-based multiple instance learning.

What carries the argument

The load-bearing object is the heterogeneity-aware optimal transport plan Q between N instance embeddings Z and K learnable survival tokens S, with a cost matrix C of Euclidean distances. Two constraints shape it: a KL divergence on the token marginal Q^T 1_N (the global long-tail constraint) and a total mass constraint 1_N^T Q 1_K = rho with rho ramped from 0.1 to 1.0 (the local uncertainty-aware constraint). A virtual survival token with a zero-cost column and an infinite KL weight absorbs the 1 - rho unselected mass, converting the problem into unbalanced optimal transport solved by a Sinkhorn-style scaling algorithm; the aggregation E = f_agg(Q^T Z) feeds a Cox hazard head.

What would settle it

Inspect the learned token marginal Q^T 1_N of a trained OTSurv model on any of the six datasets. If the marginal is close to uniform, or if substituting a genuinely long-tailed prior (for example, a Zipf distribution over token frequencies) does not change or even improves the C-index, then the specific claim that the constraint is modeling long-tailed morphology would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the core unsolved difficulty in WSI survival prediction is pathological heterogeneity at two scales — a long-tailed distribution of tissue morphologies globally, and per-tile predictive uncertainty locally — and that an optimal transport formulation is the right explicit model for both. OTSurv computes a transport plan Q from N patch embeddings to K learnable survival tokens, regularizes the token marginal with a KL constraint (called the global long-tail constraint) and the selected mass with a progressive rho constraint (called the local uncertainty-aware constraint), and aggregates Q^T Z into a survival embedding. The claim is that this plan prevents both mode collapse and excessive uniformity, selects high-confidence patches first, and yields better risk scores than previous MIL methods; the paper reports consistent C-index gains across six benchmarks and interprets the resulting attention heatmaps as aligning with tissue types.

Load-bearing premise

The argument assumes a soft KL penalty pulling the transport mass toward a uniform distribution over survival tokens preserves rare, prognostically critical patches, even though the penalty itself promotes uniformity and no measured long-tailed morphological prior is used.

Editorial extensions

If this is right

  • If OTSurv's central claim holds, WSI survival models can move from learned attention weights to an explicit transport objective whose marginal constraints encode the desired heterogeneity behavior.
  • The differentiable scaling solver means the two constraints can be trained end-to-end with standard matrix operations, making them practical on clinical slide data.
  • The progressive mass constraint offers a mechanism for automatic hard-sample selection during training, removing the need for manually tuned patch-selection thresholds.
  • The reported mean C-index of 0.646 over six TCGA cohorts would make OTSurv a strong reference point for future MIL survival methods.

Reading between the lines

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

  • The KL target in Eq. (2) is uniform over survival tokens, so the 'long-tail' regularizer literally pushes the token marginal toward equal mass; whether this preserves rare prognostically critical patches is an empirical question the paper does not directly test.
  • A testable extension is to replace the uniform KL target with a measured morphological prior from tissue-segmentation output; if C-index improves, the long-tail interpretation is literal, and if it drops, the learned uniform target is doing the work.
  • The virtual-token equivalence rests on the progressive partial transport theorem the paper cites; applying OTSurv in settings where rho is fixed or the marginal constraints are violated may still work but would not inherit that formal justification.
  • The same transport-plus-virtual-token construction could transfer to other gigapixel weakly supervised tasks, such as grade classification or mutation prediction, where rare high-signal patches dominate the label.
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

4 major / 4 minor

Summary. The paper proposes OTSurv, a multiple-instance-learning framework for whole-slide-image survival prediction that casts bag-level aggregation as an optimal transport problem. A learnable set of survival tokens is used as the target distribution, and two marginal constraints are introduced: a "global long-tail constraint" implemented as a KL penalty toward a uniform token marginal (Eq. 2), and a "local uncertainty-aware constraint" that progressively increases the total transported mass via a ramp-up parameter (Eq. 3). The resulting problem is recast as an unbalanced OT problem with a virtual token, solved by a matrix scaling algorithm. Experiments on six TCGA datasets using UNI features report a mean C-index of 0.646, an average 3.6% improvement over the best baseline, along with log-rank tests and an interpretability analysis.

Significance. The paper addresses a relevant problem in computational pathology, and the OT-based formulation is principled and clearly presented. The authors provide code, ablate the main design choices, and show that the transport plan yields interpretable heatmaps. If the empirical claim were fully supported, OTSurv would be a useful contribution to MIL-based survival prediction. However, the central "consistently outperforms" claim is not supported by the per-dataset results in Table 1, and the "long-tail" constraint is not actually long-tailed. These issues need to be resolved before the paper's main claims can be accepted.

major comments (4)
  1. [§3.2, Table 1] The statement that OTSurv "consistently outperforms" all prior methods is contradicted by the reported per-dataset C-indices. On BRCA, OTSurv (0.621±0.071) is below both PANTHER (0.650±0.139) and CLAM (0.629±0.180); on CRC, OTSurv (0.667±0.111) is below AttnMISL (0.725±0.110). The 3.6% improvement in the abstract is only a mean over six datasets, not a per-dataset win, and no paired significance test over the cross-validation folds is reported. I request paired significance tests (e.g., paired t-test or Wilcoxon on fold-wise C-indices) against each baseline, or a revised claim that does not assert consistency.
  2. [§2.2, Eq. (2)] The "global long-tail constraint" is implemented as KL(Q^T 1_N || (1/K)1_K), whose target is the uniform distribution, not a long-tailed distribution. This directly contradicts the paper's description of modeling "long-tailed morphological distributions" and "prior morphological distributions." A uniform target actively encourages equal mass across tokens and cannot be said to preserve rare, infrequent patterns without additional evidence. The authors should either define an actual long-tailed prior (e.g., a Zipf-like target) or rename the constraint and revise the abstract and Sec. 2.2 accordingly.
  3. [§3.3, Table 2] The ablation table does not support the paper's interpretation of the global constraint. The text states that "CG ✗" replaces the KL-based constraint with an equality constraint, but the table appears to label rows as "CG ✗" with varying local-constraint settings, and the row structure is ambiguous. Moreover, no ablation compares the proposed uniform-KL target against a genuinely long-tailed target or against a no-constraint variant. Without this, the claim that the constraint captures long-tailed morphology is not evidenced.
  4. [§3.1, Algorithm 1] The hyperparameter reporting is inconsistent with the algorithm. Algorithm 1 uses an entropy regularization parameter ε in M = exp(-C/ε), and the text separately states "Entropy regularization is set to λ = 0.1." Since λ is already used as the KL weight, this is ambiguous, and the value of ε is not reported. This prevents reproduction of the reported results. Please clarify the notation and report ε.
minor comments (4)
  1. [§3.2, Fig. 2(a)] The log-rank p-values report the significance of OTSurv's own median-risk split, not superiority over any baseline; the text should state this explicitly to avoid overclaiming.
  2. [§1 and §2.2] The phrase "sets new state-of-the-art results across six popular benchmarks" is repeated in the Abstract and Introduction; given Table 1, it should be replaced by a more cautious statement about average performance.
  3. [Table 1 and Table 2] There are typographical errors in the tables: "ST AD" should be "STAD," and "Epoches" should be "Epochs." The table formatting should also be cleaned so that ablation rows align with their conditions.
  4. [§3.1, Eq. (4)] The ramp-up schedule in Eq. (4) uses t, T, and I without a precise definition of the iteration index t; please specify that t is the global training iteration and note that at t=0 the mass ratio is not exactly ρ0 but ρ0 + (1-ρ0)e^{-5}.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: the OT formulation and solver are self-contained, and benchmark results are held-out measurements. The only reviewer-noted weaknesses are statistical or correctness concerns, not circular steps.

full rationale

OTSurv's derivation chain does not reduce to its inputs by construction. The method defines an explicit OT objective (Eqs. 1-3) with a global KL marginal penalty and a local total-mass constraint, then recasts it into an unbalanced OT problem (Eqs. 5-6) whose equivalence is attributed to [37], an external preprint rather than the present authors. The final risk score is trained end-to-end by minimizing the Cox partial likelihood, and all reported C-indices are obtained on held-out folds of six TCGA datasets using frozen UNI features and the data splits of [25]; no test-label quantity is fitted or renamed as a prediction. The paper's self-citations ([21], [34], [35]) are contextual related-work references for tissue heterogeneity and local uncertainty and are not load-bearing in the derivation. Two flagged weaknesses are correctness issues, not circularity: Eq. (2) penalizes the token marginal toward a uniform distribution (1/K 1_K) while calling it a 'long-tailed prior,' so the stated prior is not measured from morphology; and Table 1 shows OTSurv is not the per-dataset best on BRCA and CRC, making the '3.6% average improvement' an aggregate claim without a paired significance test. Neither weakness makes the central claim equivalent to its inputs. The SOTA assertion may be empirically overstated, but the paper is not circular.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The method is largely built on borrowed machinery: unbalanced OT and Sinkhorn scaling, the P2OT virtual-token equivalence, Cox partial likelihood, and frozen UNI features. The ledger makes explicit the free hyperparameters and the modeling assumption that a soft KL-to-uniform constraint is a long-tailed morphological prior.

free parameters (5)
  • KL weight lambda = 0.1
    Controls the global constraint strength; ablation with 0.2 gives mean C-index 0.635 vs 0.646 (Table 2), so 0.1 is chosen by data.
  • Initial mass ratio rho0 = 0.1
    Start of the curriculum ramp; 0.05 gives 0.628 and 0.2 gives 0.635, so 0.1 was selected via ablation.
  • Number of survival tokens K = 16
    Ablation K=8 and K=32 both give 0.643; 16 gives 0.646.
  • Ramp-up epochs T = 10
    T=20 and T=30 give 0.637 and 0.629, so 10 chosen.
  • Entropy regularization epsilon = not reported
    Alg. 1 uses exp(-C/epsilon), but no value is given in Sec. 3.1; the paper only lists lambda=0.1.
assumptions (5)
  • standard math KL-unbalanced OT and Sinkhorn scaling converge to the optimum
    Invoked in Sec. 2.3 and Alg. 1; foundational result cited to [13,29].
  • domain assumption Eq. (3) and Eq. (5) are equivalent via a virtual token with infinity-weighted KL
    Borrowed from P2OT [37], Sec. 2.3, not proved for this exact loss setting.
  • domain assumption Cox partial likelihood is a valid survival objective for bag-level features
    Used as the only training loss in Sec. 2.1; standard survival analysis assumption.
  • domain assumption Frozen UNI features capture morphology relevant to survival
    All patch features come from UNI [4]; if these features miss prognostic morphology, OT cannot recover it.
  • domain assumption MIL assumption of independent instances with a bag label
    Treats WSI as a bag of patches with one slide-level survival label, Sec. 2.1.
invented entities (1)
  • Virtual survival token
    purpose: Absorbs the 1-rho unselected transport mass so Eq. (3) becomes an unbalanced OT problem solvable by scaling algorithms.
    Introduced in Sec. 2.3 and Fig. 1(b); it is a mathematical device with no tissue or clinical counterpart, and no falsifiable prediction outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OTSurv: A Novel Multiple Instance Learning Framework for Survival Prediction with Heterogeneity-aware Optimal Transport." pith.science (2026). https://pith.science/paper/RFVIBODZ

@misc{pith2026250620741,
  author       = {Pith},
  title        = {Pith review of: OTSurv: A Novel Multiple Instance Learning Framework for Survival Prediction with Heterogeneity-aware Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RFVIBODZ}},
  note         = {Machine review of arXiv:2506.20741}
}
read the original abstract

Survival prediction using whole slide images (WSIs) can be formulated as a multiple instance learning (MIL) problem. However, existing MIL methods often fail to explicitly capture pathological heterogeneity within WSIs, both globally -- through long-tailed morphological distributions, and locally through -- tile-level prediction uncertainty. Optimal transport (OT) provides a principled way of modeling such heterogeneity by incorporating marginal distribution constraints. Building on this insight, we propose OTSurv, a novel MIL framework from an optimal transport perspective. Specifically, OTSurv formulates survival predictions as a heterogeneity-aware OT problem with two constraints: (1) global long-tail constraint that models prior morphological distributions to avert both mode collapse and excessive uniformity by regulating transport mass allocation, and (2) local uncertainty-aware constraint that prioritizes high-confidence patches while suppressing noise by progressively raising the total transport mass. We then recast the initial OT problem, augmented by these constraints, into an unbalanced OT formulation that can be solved with an efficient, hardware-friendly matrix scaling algorithm. Empirically, OTSurv sets new state-of-the-art results across six popular benchmarks, achieving an absolute 3.6% improvement in average C-index. In addition, OTSurv achieves statistical significance in log-rank tests and offers high interpretability, making it a powerful tool for survival prediction in digital pathology. Our codes are available at https://github.com/Y-Research-SBU/OTSurv.

Figures

Figures reproduced from arXiv: 2506.20741 by the authors.

Figure 1
Figure 1. Overview of the OTSurv framework. (a) Framework of OTSurv: A WSI W is processed into patch features Z, aligned with survival tokens S via heterogeneity-aware OT, and aggregated into a survival embedding E. (b) heterogeneity-aware Optimal Transport: By extending the cost matrix with an additional column of zeros (i.e., virtual token), the unbalanced OT solver can be used for solving heterogeneity-aware OT with global… view at source ↗
Figure 2
Figure 2. (a) Kaplan–Meier survival curves, (b) OTSurv interpretability analysis. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [37]

    arXiv preprint arXiv:2401.09266 (2024)

    Zhang, C., Ren, H., He, X.: P2 ot: Progressive partial optimal transport for deep imbalanced clustering. arXiv preprint arXiv:2401.09266 (2024)

  2. [1]

    BMJ328(7447), 1073 (2004)

    Bland, J.M., Altman, D.G.: The logrank test. BMJ328(7447), 1073 (2004)

  3. [2]

    Nature Medicine (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 (2019)

  4. [3]

    In: CVPR (2022)

    Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: CVPR (2022)

  5. [4]

    Nature Medicine (2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Song, A.H., Chen, B., Zhang, A., Shao, D., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024)

  6. [5]

    Journal of the Royal Statistical So- ciety: Series B (Methodological) (1972)

    Cox, D.R.: Regression models and life-tables. Journal of the Royal Statistical So- ciety: Series B (Methodological) (1972)

  7. [6]

    New England Journal of Medicine (2012)

    Gerlinger, M., Rowan, A.J., Horswell, S., Larkin, J., Endesfelder, D., Gronroos, E., Martinez, P., Matthews, N., Stewart, A., Tarpey, P., et al.: Intratumor hetero- geneity and branched evolution revealed by multiregion sequencing. New England Journal of Medicine (2012)

  8. [7]

    In: CVPR (2016)

    Hou, L., Samaras, D., Kurc, T.M., Gao, Y., Davis, J.E., Saltz, J.H.: Patch-based convolutional neural network for whole slide tissue image classification. In: CVPR (2016)

Show all 38 references
  1. [8]

    Nature Communications (2021)

    Howard, F.M., Dolezal, J., Kochanny, S., Schulte, J., Chen, H., Heij, L., Huo, D., Nanda, R., Olopade, O.I., Kather, J.N., et al.: The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nature Communications (2021)

  2. [9]

    In: ICML (2018)

    Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: ICML (2018)

  3. [10]

    IEEE Transactions on Medical Imaging (2024)

    Jin, C., Luo, L., Lin, H., Hou, J., Chen, H.: Hmil: Hierarchical multi-instance learning for fine-grained whole slide image classification. IEEE Transactions on Medical Imaging (2024)

  4. [11]

    In: Doklady Akademii Nauk

    Kantorovich, L.: On the transfer of masses (in russian). In: Doklady Akademii Nauk. vol. 37, p. 227 (1942) 10 Q. Ren et al

  5. [12]

    IEEE Transac- tions on Pattern Analysis and Machine Intelligence (2024)

    Khamis, A., Tsuchida, R., Tarek, M., Rolland, V., Petersson, L.: Scalable optimal transport methods in machine learning: A contemporary survey. IEEE Transac- tions on Pattern Analysis and Machine Intelligence (2024)

  6. [13]

    SIAM Journal on Matrix Analysis and Applications (2008)

    Knight, P.A.: The sinkhorn–knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications (2008)

  7. [14]

    In: CVPR (2023)

    Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task- specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: CVPR (2023)

  8. [15]

    Inventiones Mathemat- icae (2018)

    Liero, M., Mielke, A., Savaré, G.: Optimal entropy-transport problems and a new hellinger–kantorovich distance between positive measures. Inventiones Mathemat- icae (2018)

  9. [16]

    In: CVPR (2023)

    Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: CVPR (2023)

  10. [17]

    Cell (2018)

    Liu, J., Lichtenberg, T., Hoadley, K.A., Poisson, L.M., Lazar, A.J., Cherniack, A.D., Kovatich, A.J., Benz, C.C., Levine, D.A., Lee, A.V., et al.: An integrated tcga pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell (2018)

  11. [18]

    Nature Biomedical Engineering (2021)

    Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering (2021)

  12. [19]

    arXiv preprint arXiv:2505.20202 (2025)

    Ma, J., Xu, Y., Zhou, F., Wang, Y., Jin, C., Guo, Z., Wu, J., Tang, O.K., Zhou, H., Wang, X., et al.: Pathbench: A comprehensive comparison bench- mark for pathology foundation models towards precision oncology. arXiv preprint arXiv:2505.20202 (2025)

  13. [20]

    Cell (2017)

    McGranahan, N., Swanton, C.: Clonal heterogeneity and tumor evolution: past, present, and the future. Cell (2017)

  14. [21]

    In: MICCAI (2023)

    Ren, Q., Zhao, Y., He, B., Wu, B., Mai, S., Xu, F., Huang, Y., He, Y., Huang, J., Yao, J.: Iib-mil: Integrated instance-level and bag-level multiple instances learning with label disambiguation for pathological image analysis. In: MICCAI (2023)

  15. [22]

    In: ICLR (2017)

    Samuli, L., Timo, A.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)

  16. [23]

    In: NeurIPS (2021)

    Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Trans- former based correlated multiple instance learning for whole slide image classifica- tion. In: NeurIPS (2021)

  17. [24]

    In: CVPR (2024)

    Song, A.H., Chen, R.J., Ding, T., Williamson, D.F., Jaume, G., Mahmood, F.: Morphological prototyping for unsupervised slide representation learning in com- putational pathology. In: CVPR (2024)

  18. [25]

    In: ICML (2024)

    Song, A.H., Chen, R.J., Jaume, G., Vaidya, A.J., Baras, A., Mahmood, F.: Multi- modal prototyping for cancer survival prediction. In: ICML (2024)

  19. [26]

    Cell (2024)

    Song, A.H., Williams, M., Williamson, D.F., Chow, S.S., Jaume, G., Gao, G., Zhang, A., Chen, B., Baras, A.S., Serafin, R., et al.: Analysis of 3d pathology samples using weakly supervised ai. Cell (2024)

  20. [27]

    In: ICCV (2023)

    Tang, W., Huang, S., Zhang, X., Zhou, F., Zhang, Y., Liu, B.: Multiple instance learning framework with masked hard instance mining for whole slide image clas- sification. In: ICCV (2023)

  21. [28]

    NeurIPS (2017)

    Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS (2017)

  22. [29]

    Villani, C., et al.: Optimal transport: old and new, vol. 338. Springer (2009)

  23. [30]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) Title Suppressed Due to Excessive Length 11

    Wang, X., Chen, Y., Zhu, W.: A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) Title Suppressed Due to Excessive Length 11

  24. [31]

    In: ICLR (2023)

    Xiang, J., Zhang, J.: Exploring low-rank property in multiple instance learning for whole slide image classification. In: ICLR (2023)

  25. [32]

    Nature (2024)

    Xu, H., Usuyama, N., Bagga, J., Zhang, S., Rao, R., Naumann, T., Wong, C., Gero, Z., González, J., Gu, Y., et al.: A whole-slide foundation model for digital pathology from real-world data. Nature (2024)

  26. [33]

    Medical Image Analysis (2020)

    Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis (2020)

  27. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    You, C., Dai, W., Liu, F., Min, Y., Dvornek, N.C., Li, X., Clifton, D.A., Staib, L., Duncan, J.S.: Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

  28. [35]

    In: NeurIPS (2023)

    You, C., Dai, W., Min, Y., Liu, F., Clifton, D., Zhou, S.K., Staib, L., Duncan, J.: Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. In: NeurIPS (2023)

  29. [36]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)

    Zadeh, S.G., Schmid, M.: Bias in cross-entropy-based training of deep survival networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)

  30. [38]

    In: CVPR (2022)

    Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathol- ogy whole slide image classification. In: CVPR (2022)

Pith tools

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