Pith. sign in

REVIEW 5 major objections 5 minor 53 references

Dual-stream attention-guided learning for weakly supervised whole slide image classification

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DSAGL aims to establish that a dual-stream teacher-student architecture, converting bag-level attention scores into soft instance pseudo-labels, beats existing MIL methods at both instance and bag levels on three benchmarks.

desk verdict Competent incremental MIL architecture whose central 'consistent superiority' claim is refuted by its own Table 1 at the 1% positive-ratio setting. read the letter →

arxiv 2505.23341 v3 pith:E7BBEDQJ submitted 2025-05-29 cs.CV

Daoxi Cao , Hangbei Cheng , Yijin Li , Ruolin Zhou , Xuehan Zhang , Xinyi Li , Binwei Li , Xuancheng Gu
show 3 more authors
Jianan Zhang Xueyu Liu Yongfei Wu
This is my paper
classification cs.CV
keywords wholeslideimageclassificationmultipleinstancelearningweaklysupervisedteacher-studentnetworkattentionmechanismpseudo-labeldistillationMambaencodercomputationalpathology
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 tries to establish that weakly supervised whole-slide image classification can be improved by decoupling bag-level and instance-level learning into a teacher-student pair. The teacher classifies the slide and, in doing so, produces per-patch attention scores; those scores are normalized into soft pseudo-labels that train the student to classify individual patches. The authors argue that this attention-guided pseudo-labeling, an alternating training schedule, a linear-complexity Mamba encoder, and a multi-scale attention module together resolve the usual trade-off between instance-level discriminability and bag-level accuracy. On synthetic bags from CIFAR-10 and NCT-CRC and on real TCGA-Lung slides, they report instance-level AUCs up to 0.9287 and 0.9916 and a bag-level AUC of 0.9727, consistently above recent MIL baselines. If true, the contribution is a way to get patch-level localization from slide-level labels alone, which matters in pathology where fine-grained annotations are costly.

What carries the argument

The load-bearing machinery is the dual-stream teacher-student loop. The teacher branch is built around FASA, a fusion-attentive scale-aware module that runs parallel $3\times3$, $5\times5$, and $7\times7$ convolutions, applies channel and spatial attention, and pools the result into bag-level logits plus per-instance attention scores. NormProb converts those scores into soft pseudo-labels $\hat{z}_{i,j}$ used as student supervision. The student branch has its own classification head and is trained with a hybrid loss, weighted cross-entropy plus a temperature-scaled KL term, so it learns from both hard and soft labels. VSSMamba, a skip-connected selective Mamba encoder with linear complexity, supplies the shared instance features, and alternating updates between teacher and student let the pseudo-labels improve as training progresses.

What would settle it

Construct synthetic bags in which the known positive instances are deliberately made visually bland while negative instances carry the bag-level surface statistics; if the student's top-ranked instances do not match the known positives, the attention pseudo-labels are encoding bag composition rather than diagnostic patch-level signal.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that joint instance- and bag-level performance under weak supervision is achievable through a dual-stream architecture that converts teacher attention into student supervision. The teacher stream, built around the FASA module, aggregates instance features into a bag prediction and emits an attention score per patch; NormProb turns these scores into probabilistic pseudo-labels $\hat{z}_{i,j}$. The student stream is trained to predict per-instance classes from the same shared VSSMamba features using a hybrid loss that blends weighted cross-entropy on hard pseudo-labels with KL divergence against the teacher's soft outputs. The two streams are updated alternately rather than jointly or strictly sequentially, so that pseudo-labels can refine during training. The paper reports that this configuration outperforms MIL baselines including IBMIL and R2T-MIL, reaching instance-level AUCs of 0.9287 on CIFAR-10 and 0.9916 on NCT-CRC, and a bag-level AUC of 0.9727 on TCGA-Lung.

Load-bearing premise

The approach stands or falls on whether the teacher's per-patch attention scores, normalized into soft pseudo-labels, point at the truly diagnostic patches; if they merely reflect which bag a patch came from, the reported instance-level gains would be an artifact of bag composition.

Editorial extensions

If this is right

  • Within the paper's claims, attention-derived pseudo-labels can substitute for missing patch annotations, so the same slide-level labels yield both a whole-slide diagnosis and a patch-level heatmap.
  • The alternating teacher-student update is presented as a stability mechanism: it should reduce early-training pseudo-label noise and let instance-level and bag-level objectives converge together rather than interfere.
  • Because VSSMamba processes long patch sequences in linear time, the approach should scale to WSIs with hundreds of thousands of patches without the quadratic cost of Transformer attention.
  • If the reported numbers hold, DSAGL would set new reference points on CIFAR-10, NCT-CRC, and TCGA-Lung for weakly supervised MIL, at both instance and bag levels.

Reading between the lines

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

  • The ablation in which pseudo-label training (0.9438 instance AUC) slightly beats ground-truth instance labels (0.9420) suggests the student may be absorbing bag-composition information; a test set that controls for positive-instance ratio across bags would show whether instance-level gains reflect true localization or bag-level shortcuts.
  • The same teacher-student pseudo-label loop could transfer to other weakly supervised problems with coarse group labels, such as video event detection or remote-sensing scene classification, where per-item labels are expensive.
  • Explicitly modeling pseudo-label uncertainty, which the paper lists as a limitation, is a natural next step; calibrated or Bayesian attention outputs might reduce early-stage noise in sparse-lesion cases.
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 DSAGL, a dual-stream teacher-student framework for weakly supervised whole slide image classification. A teacher branch uses a Mamba-based encoder (VSSMamba) and a multi-scale attention module (FASA) to produce bag-level predictions and attention scores; these scores are normalized into soft pseudo-labels that supervise a student branch for instance-level classification. An alternating training strategy and a hybrid loss combining weighted cross-entropy and KL divergence are introduced. Experiments are reported on CIFAR-10, NCT-CRC, and TCGA-Lung, with the central claim that DSAGL consistently outperforms state-of-the-art MIL methods at both instance and bag levels, especially under low positive-instance ratios.

Significance. If the claims were substantiated, an efficient Mamba-based dual-stream MIL framework with attention-guided pseudo-labeling would be a useful contribution to computational pathology, particularly for instance-level localization under slide-level supervision. The paper identifies a real problem (weakly supervised instance discrimination in WSIs) and combines several plausible components. However, the experimental evidence as presented does not support the headline claim of consistent superiority: the paper's own Table 1 contains a direct counterexample at the 1% positive ratio, several experimental-protocol details are missing, and the pseudo-label/ground-truth comparison in Table 5 is internally anomalous. The strengths of the work—the architectural idea and the breadth of datasets—cannot compensate for the fact that the central quantitative claim is contradicted by the reported results.

major comments (5)
  1. [Table 1(a), Section 4.4.1] The central claim that DSAGL 'consistently outperforms state-of-the-art MIL methods' is directly contradicted by the CIFAR-10 results at the 1% positive-patch ratio. DSAGL's instance-level AUC is 0.5971, which is below MI-Net (0.6447), ABMIL (0.6225), CLAM (0.6054), TransMIL (0.7015), IBMIL (0.7353), and R2T-MIL (0.6714). At the bag level in Table 1(b), DSAGL's 0.6044 is also below R2T-MIL's 0.6874. The 1% setting is explicitly within the paper's experimental scope, and the conclusion specifically claims strength under low positive-instance ratios. This internal counterexample means the headline claim cannot stand as written without additional qualification or revised experiments.
  2. [Sections 4.1.2, 4.2, Eq. (20)] Several load-bearing experimental and algorithmic details are missing. First, the synthetic bag construction is not specified: for NCT-CRC, it is not stated which of the nine tissue classes serve as positive instances, what the bag size is, how many bags are generated, or how train/test splits are performed; for CIFAR-10, the rule for combining classes into positive and negative bags is not given. Second, NormProb in Eq. (20) is never defined, so the conversion of attention scores into pseudo-labels is not reproducible. Third, the values of alpha, w_n, T, and epsilon in the hybrid loss (Eqs. 23-26) are not reported, and the number of Mamba layers n is only given as 0, 2, or 3 per dataset without any ablation or sensitivity analysis. Without these details, the reported instance-level AUCs cannot be interpreted or compared fairly.
  3. [Section 4.3, Table 3] The TCGA-Lung comparison is not a controlled experiment. The text states that 'Baseline results are either reproduced under our unified experimental settings or cited directly from existing studies on weakly supervised WSI classification [36].' Table 3 appears to rely on cited baselines, but no information is provided about the feature extractor, patch sampling, or evaluation protocol used for those cited numbers. The reported advantage of DSAGL over R2T-MIL is only 0.0026 AUC (0.9727 vs. 0.9701) and over IBMIL is 0.0038, so protocol differences could easily account for the difference. The authors should rerun all baselines under the same pipeline or clearly mark which entries are from prior work and justify comparability.
  4. [Table 5, Section 4.5.2] The pseudo-label ablation contains an internally inconsistent 'upper bound.' The row labeled 'GT'—ground-truth instance labels—reports an instance-level AUC of 0.9420, while the full configuration with soft pseudo-labels reaches 0.9438. If the pseudo-labels are derived solely from bag-level attention scores, they should not provide more instance-level information than true instance labels. The paper dismisses this as 'slightly lower' and claims pseudo-labeling 'approaches' fully supervised performance, but the ordering actually suggests that the comparison is confounded by configuration differences (the GT row does not use KL divergence) or that the instance-level metric is being inflated by bag-composition artifacts. This needs a clear explanation or a corrected experiment.
  5. [Tables 1-3] No error bars, standard deviations, or repeated-seed results are reported for any of the main tables. Several differences highlighted as improvements are small (e.g., TCGA-Lung bag-level AUC 0.9727 vs. 0.9701), and the NCT-CRC baseline values are non-monotonic in ways that suggest instability (e.g., CLAM instance-level AUC 0.8525 at 20% but 0.5450 at 50%). Without multiple seeds or statistical testing, the claimed consistent superiority could be a single-run artifact. The authors should report mean and standard deviation over at least three independent runs, especially for the CIFAR-10 and NCT-CRC synthetic settings where computational cost is manageable.
minor comments (5)
  1. [Section 3.4 title] The heading 'Duel-Stream Network' is a typo; it should be 'Dual-Stream Network.'
  2. [References] Reference [20] is cited in Section 2.2 as Co-MIL, but the full reference is a paper on Fe/Co-MIL-88(NH2) metal-organic frameworks and extracellular vesicle detection; this is clearly the wrong citation and should be replaced with the actual Co-MIL source.
  3. [Section 4.1.3] The text claims that at 20x magnification 'each WSI is partitioned into approximately 5.2 million image patches' of size 224x224. This is implausible for a single WSI; the number likely refers to the total across the 1,054 WSIs. Please clarify.
  4. [Section 4.3 vs. Section 4.4.1] Section 4.3 says baseline results are either reproduced or cited from existing studies, while Section 4.4.1 and Fig. 6 state that 'all baseline models are reimplemented using publicly available code under identical settings.' These statements are inconsistent and should be reconciled.
  5. [Eq. (25)] The direction of the KL divergence in Eq. (25) is ambiguous. The text says q and p are the student's and teacher's predicted probabilities, but standard knowledge distillation typically uses the teacher distribution as the target. Please clarify the roles of q and p explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity by construction: the instance-level evaluation is external, and the teacher-student pseudo-label loop is not an equation-level reduction.

full rationale

DSAGL's derivation chain is not circular under the strict definition. The teacher stream (Eqs. 17-19) maps bag features to attention scores and a bag logit supervised by bag labels; the student stream (Eqs. 20-22) is trained to match NormProb-normalized teacher attention via a hybrid CE+KL loss. Although the pseudo-labels are generated by the model itself, there is no equation in which the reported instance predictions are defined to equal those pseudo-labels: the student's head P_s is optimized, not set equal, to the pseudo-label, and the instance-level AUC used in Tables 1 and 2 is computed against ground-truth instance labels, an external benchmark. The Table 5 observation that soft pseudo-labels (0.9438) slightly exceed ground-truth instance labels (0.9420) is an empirical anomaly, not a construction-level equivalence, and is more a correctness and interpretability concern about what the attention scores encode. The paper's central claim of consistent superiority is contradicted by its own Table 1(a) at the 1% positive ratio on CIFAR-10 (DSAGL 0.5971 vs IBMIL 0.7353), but a false or unsupported empirical claim is a correctness risk, not circularity. The only self-citation, GLMKD [6], appears in related work and is not load-bearing for any derivation. No fitted quantity is renamed as a prediction, no uniqueness theorem is imported from the authors, and no known result is merely renamed. Therefore no specific circular step can be exhibited.

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

The central claim rests on a handful of hand-chosen hyperparameters (alpha, w_n, T, epsilon) and on the assumption that teacher attention yields valid instance pseudo-labels. The latter is directly contradicted by the Table 5 anomaly, making it the most fragile input.

free parameters (5)
  • alpha (Eq. 26) = not reported
    Weight balancing CE and KL student losses; value tuned but undisclosed.
  • w_n (Eq. 24) = not reported
    Negative-instance weight in student CE loss; value tuned but undisclosed.
  • T (Eq. 25) = not reported
    Distillation temperature, stated T>1; exact value undisclosed.
  • epsilon (Eq. 23) = not reported
    Numerical stability constant in teacher BCE loss.
  • Number of Mamba layers n = 0 (CIFAR-10), 2 (NCT-CRC), 3 (TCGA-Lung)
    Handpicked per dataset based on resolution; part of architecture search.
assumptions (5)
  • domain assumption Standard MIL assumption: a bag is positive iff at least one instance is positive (Eq. 1).
    The entire problem formulation relies on this definition; in real WSIs, the relationship between patch-level and slide-level labels is more complex.
  • ad hoc to paper Teacher attention scores, after NormProb normalization, are usable as instance-level pseudo-labels (Eq. 20).
    No calibration or validation shows attention scores are calibrated probabilities; Table 5 suggests they encode bag-level info.
  • domain assumption Synthetic bags built from CIFAR-10 and NCT-CRC patches faithfully emulate WSI diagnostic structure.
    Randomly combining images from different classes is a crude proxy for spatially coherent tumor regions in real slides; positive class selection is not specified.
  • ad hoc to paper Baselines cited from prior studies are directly comparable under the same evaluation protocol.
    For TCGA-Lung, results are cited from ref [36] with potentially different splits, feature extractors, and training procedures.
  • domain assumption The Mamba-based encoder models long-range dependencies among patches with linear complexity, sufficient for WSI classification.
    This is the motivation for VSSMamba; no analysis beyond empirical results supports the sufficiency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-stream attention-guided learning for weakly supervised whole slide image classification." pith.science (2026). https://pith.science/paper/E7BBEDQJ

@misc{pith2026250523341,
  author       = {Pith},
  title        = {Pith review of: Dual-stream attention-guided learning for weakly supervised whole slide image classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7BBEDQJ}},
  note         = {Machine review of arXiv:2505.23341}
}
read the original abstract

Whole slide images (WSIs) play a crucial role in cancer diagnosis due to their ultra-high resolution and rich morphological information, and multiple instance learning (MIL) has become a prevalent paradigm to solve the massive size of WSIs and the scarcity of fine-grained annotations of instance. However, most existing MIL methods struggle to accurately identify diagnostically critical local regions (instance) using only slide-level labels, and suffer from modelling the relationship of instances efficiently. To address these defects, we propose a Dual-Stream Attention-Guided Learning (DSAGL) framework. DSAGL bridges slide-level supervision and instance-level learning through a teacher-student dual-stream architecture, and mitigates instance ambiguity by generating attention-guided pseudo labels. The framework employs a shared lightweight encoder to efficiently model long-range dependencies and an attention-based fusion mechanism to enhance sensitivity to sparse, informative regions. Extensive experiments on synthetic benchmarks and real-world pathological WSI datasets demonstrate that DSAGL consistently outperforms state-of-the-art MIL methods, achieving superior discriminative performance and robustness under weak supervision.

Figures

Figures reproduced from arXiv: 2505.23341 by the authors.

Figure 1
Figure 1. (a) Basic MIL architecture; (b) Our proposed DSAGL architecture. D.Cao et al.: Preprint submitted to Biomedical Signal Processing and Control Page 2 of 20 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (b) Supervised learning and Multiple-Instance Learning (MIL). Supervised learning and Multiple-Instance Learning(MIL) 3.1.2. SSM and Mamba Structured State Space Models (SSMs) [52] offer an efficient module for long-range sequence modeling with linear time complexity. Given a 1D input sequence 𝑥(𝑡), the system evolves through a hidden state ℎ(𝑡) ∈ ℝN , where N denotes the dimension of the hidden state, governed by t… view at source ↗
Figure 3
Figure 3. (a) The overall architecture of the proposed Dual-Stream Attention-Guided Learning Weakly Supervised Framework (DSAGL); (b) The proposed Vision-aware Skip-connected Selective Mamba encoder (VSSMamba); (c) The proposed Fusion-Attentive Scale-Aware module (FASA). To enable efficient and interpretable pathological image classification, we propose a Dual-Stream Attention-Guided Learning framework (DSAGL) for weakly supe… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Examples of the three datasets used in our experiments. (a) CIFAR-10, consisting of 10 categories of natural images; (b) NCT-CRC, consisting of 9 categories of colorectal histopathology image patches; (c) TCGA-Lung, consisting of whole-slide images from two lung cancer…
Figure 5
Figure 5. Figure 5: Comparison of AUC for Different Classification Levels on the CIFAR-10 dataset. strong generalization capability across heterogeneous tissue patterns and its ability to retain discriminative precision under weak supervision. 4.4.3. Results on the TCGA-Lung dataset For t…
Figure 6
Figure 6. Figure 6: Test AUC curves on the CIFAR-10 dataset [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Comparison of AUC for Different Classification Levels on the NCT-CRC dataset [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of attention heatmaps on four tissue classes (ADI, MUS, LYM, TUM) from the NCT￾CRC dataset. Each row corresponds to a representative MIL model: DSAGL (Ours), R2T-MIL (Transformer-based), CLAM (Attention-based), and RNN-MIL (Classical pooling-base…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages

  1. [36]

    Qu,L.,Wang,M.,Song,Z.,etal.,2022.Bi-directionalweaklysupervisedknowledgedistillationforwholeslideimageclassification.Advances in Neural Information Processing Systems 35, 15368–15381

    work page 2022

  2. [1]

    Alex, K., Sutskever, I., Hinton, G.E., 2011. Imagenet classification with deep convolutional networks, in: volume-1; pages-1097–1105; NIPS’12 Proceedings of the 25th International Conference on Neural Information Processing System

    work page 2011

  3. [2]

    Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., Santamaría, J., Fadhel, M.A., Al-Amidie, M., Farhan, L.,

    work page

  4. [3]

    Multiple instance classification: Review, taxonomy and comparative study

    Amores, J., 2013. Multiple instance classification: Review, taxonomy and comparative study. Artificial intelligence 201, 81–105

    work page 2013

  5. [4]

    Clinical-grade computational pathology using weakly supervised deep learning on whole slide images

    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., 2019. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine 25, 1301–1309

    work page 2019

  6. [5]

    MSA-MIL: A deep residual multiple instance learning model based on multi-scale annotation for classification and visualization of glomerular spikes

    Chen, Y., Li, M., Wu, Y., Liu, X., Hao, F., Zhou, D., Zhou, X., Wang, C., 2020. Msa-mil: A deep residual multiple instance learning model based on multi-scale annotation for classification and visualization of glomerular spikes. arXiv preprint arXiv:2007.00858

    work page Pith review arXiv 2020

  7. [6]

    Glmkd: Joint global and local mutual knowledge distillation for weakly supervised lesion segmentation in histopathology images

    Cheng, H., Liu, X., Zhang, J., Dong, X., Ma, X., Zhang, Y., Meng, H., Chen, X., Yue, G., Li, Y., et al., 2025. Glmkd: Joint global and local mutual knowledge distillation for weakly supervised lesion segmentation in histopathology images. Expert Systems with Applications 279, 127425

    work page 2025

  8. [7]

    The cnn paradigm

    Chua, L.O., Roska, T., 1993. The cnn paradigm. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications 40, 147–156

    work page 1993

Show all 53 references
  1. [8]

    Deep learning for whole slide image analysis: an overview

    Dimitriou, N., Arandjelović, O., Caie, P.D., 2019. Deep learning for whole slide image analysis: an overview. Frontiers in medicine 6, 264

    2019

  2. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale

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

    2020 arXiv

  3. [10]

    Du, F., Yang, P., Jia, Q., Nan, F., Chen, X., Yang, Y., 2023. Global and local mixture consistency cumulative learning for long-tailed visual recognitions, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 15814–15823

    2023

  4. [11]

    Whole slide imaging in pathology: advantages, limitations, and emerging perspectives

    Farahani, N., Parwani, A.V., Pantanowitz, L., 2015. Whole slide imaging in pathology: advantages, limitations, and emerging perspectives. Pathology and Laboratory Medicine International , 23–33

    2015

  5. [12]

    Mamba: Linear-time sequence modeling with selective state spaces

    Gu, A., Dao, T., 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    2023 arXiv

  6. [13]

    Efficiently modeling long sequences with structured state spaces

    Gu, A., Goel, K., Ré, C., 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396

    2021 arXiv

  7. [14]

    Asurveyonvisiontransformer

    Han,K.,Wang,Y.,Chen,H.,Chen,X.,Guo,J.,Liu,Z.,Tang,Y.,Xiao,A.,Xu,C.,Xu,Y.,etal.,2022. Asurveyonvisiontransformer. IEEE transactions on pattern analysis and machine intelligence 45, 87–110

    2022

  8. [15]

    Transformer in transformer

    Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y., 2021. Transformer in transformer. Advances in neural information processing systems 34, 15908–15919

    2021

  9. [16]

    Themeaninganduseoftheareaunderareceiveroperatingcharacteristic(roc)curve

    Hanley,J.A.,McNeil,B.J.,1982. Themeaninganduseoftheareaunderareceiveroperatingcharacteristic(roc)curve. Radiology143,29–36

    1982

  10. [17]

    T-mamba:Frequency-enhancedgatedlong-rangedependencyfortooth3dcbctsegmentation

    Hao,J.,He,L.,Hung,K.F.,2024. T-mamba:Frequency-enhancedgatedlong-rangedependencyfortooth3dcbctsegmentation. arXive-prints , arXiv–2404

    2024

  11. [18]

    Hou, W., Yu, L., Lin, C., Huang, H., Yu, R., Qin, J., Wang, L., 2022. Hˆ 2-mil: exploring hierarchical representation with heterogeneous multiple instance learning for whole slide image analysis, in: Proceedings of the AAAI conference on artificial intelligence, pp. 933–941

    2022

  12. [19]

    Attention-based deep multiple instance learning, in: International conference on machine learning, PMLR

    Ilse, M., Tomczak, J., Welling, M., 2018. Attention-based deep multiple instance learning, in: International conference on machine learning, PMLR. pp. 2127–2136

    2018

  13. [20]

    Bimetallic metal–organic framework fe/co-mil-88 (nh2) exhibiting high peroxidase-like activity and its application in detection of extracellular vesicles

    Jiang, Q., Xiao, Y., Hong, A.N., Gao, Z., Shen, Y., Fan, Q., Feng, P., Zhong, W., 2022. Bimetallic metal–organic framework fe/co-mil-88 (nh2) exhibiting high peroxidase-like activity and its application in detection of extracellular vesicles. ACS Applied Materials & Interfaces...

    2022

  14. [21]

    Multi-class texture analysis in colorectal cancer histology

    Kather, J.N., Weis, C.A., Bianconi, F., Melchers, S.M., Schad, L.R., Gaiser, T., Marx, A., Zöllner, F.G., 2016. Multi-class texture analysis in colorectal cancer histology. Scientific reports 6, 1–11

    2016

  15. [22]

    Adam: A method for stochastic optimization

    Kingma, D.P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  16. [23]

    Learning multiple layers of features from tiny images

    Krizhevsky, A., Hinton, G., et al., 2009. Learning multiple layers of features from tiny images

    2009

  17. [24]

    On information and sufficiency

    Kullback, S., Leibler, R.A., 1951. On information and sufficiency. The annals of mathematical statistics 22, 79–86

    1951

  18. [25]

    Backpropagation applied to handwritten zip code recognition

    LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W., Jackel, L.D., 1989. Backpropagation applied to handwritten zip code recognition. Neural computation 1, 541–551

    1989

  19. [26]

    Li, B., Li, Y., Eliceiri, K.W., 2021. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14318–14328

    2021

  20. [27]

    Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L., 2023a. Task-specific fine-tuning via variational information bottleneckforweakly-supervisedpathologywholeslideimageclassification,in:ProceedingsoftheIEEE/CVFconferenceoncomputervision and pattern re...

  21. [28]

    Weakly supervised histopathology image segmentation with self-attention

    Li, K., Qian, Z., Han, Y., Chang, E.I.C., Wei, B., Lai, M., Liao, J., Fan, Y., Xu, Y., 2023b. Weakly supervised histopathology image segmentation with self-attention. Medical Image Analysis 86, 102791

  22. [29]

    Interventional bag multi-instance learning on whole-slide pathological images, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W., 2023. Interventional bag multi-instance learning on whole-slide pathological images, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19830–19839

    2023

  23. [30]

    Video swin transformer, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H., 2022. Video swin transformer, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3202–3211

    2022

  24. [31]

    Data-efficient and weakly supervised computational pathology on whole-slide images

    Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F., 2021. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5, 555–570

    2021

  25. [32]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Ma, J., Li, F., Wang, B., 2024. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722

    2024 arXiv

  26. [33]

    Cross-entropy loss functions: Theoretical analysis and applications, in: International conference on Machine learning, PMLR

    Mao, A., Mohri, M., Zhong, Y., 2023. Cross-entropy loss functions: Theoretical analysis and applications, in: International conference on Machine learning, PMLR. pp. 23803–23828. D.Cao et al.:Preprint submitted to Biomedical Signal Processing and Control Page 19 of 20 Weakly S...

    2023

  27. [34]

    Nakhli, R., Moghadam, P.A., Mi, H., Farahani, H., Baras, A., Gilks, B., Bashashati, A., 2023. Sparse multi-modal graph transformer with shared-contextprocessingforrepresentationlearningofgiga-pixelimages,in:ProceedingsoftheIEEE/CVFConferenceonComputerVision and Pattern Recogni...

    2023

  28. [35]

    Digital pathology and artificial intelligence

    Niazi, M.K.K., Parwani, A.V., Gurcan, M.N., 2019. Digital pathology and artificial intelligence. The lancet oncology 20, e253–e261

    2019

  29. [37]

    A stochastic approximation method

    Robbins, H., Monro, S., 1951. A stochastic approximation method. The annals of mathematical statistics , 400–407

    1951

  30. [38]

    Ocelot: overlapped cell on tissuedatasetforhistopathology,in:ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition,pp.23902–23912

    Ryu, J., Puche, A.V., Shin, J., Park, S., Brattoli, B., Lee, J., Jung, W., Cho, S.I., Paeng, K., Ock, C.Y., et al., 2023. Ocelot: overlapped cell on tissuedatasetforhistopathology,in:ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition,pp.23902–23912

    2023

  31. [39]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification

    Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al., 2021. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems 34, 2136–2147

    2021

  32. [40]

    Shi, J., Li, C., Gong, T., Zheng, Y., Fu, H., 2024. Vila-mil: Dual-scale vision-language multiple instance learning for whole slide image classification, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11248–11258

    2024

  33. [41]

    Astructure-awarehierarchicalgraph-basedmultiple instance learning framework for pt staging in histopathological image

    Shi,J.,Tang,L.,Li,Y.,Zhang,X.,Gao,Z.,Zheng,Y.,Wang,C.,Gong,T.,Li,C.,2023. Astructure-awarehierarchicalgraph-basedmultiple instance learning framework for pt staging in histopathological image. IEEE Transactions on Medical Imaging 42, 3000–3011

    2023

  34. [42]

    Artificial intelligence in histopathology: enhancing cancer research and clinical oncology

    Shmatko, A., Ghaffari Laleh, N., Gerstung, M., Kather, J.N., 2022. Artificial intelligence in histopathology: enhancing cancer research and clinical oncology. Nature cancer 3, 1026–1038

    2022

  35. [43]

    Artificial intelligence for digital and computational pathology

    Song, A.H., Jaume, G., Williamson, D.F., Lu, M.Y., Vaidya, A., Miller, T.R., Mahmood, F., 2023. Artificial intelligence for digital and computational pathology. Nature Reviews Bioengineering 1, 930–949

    2023

  36. [44]

    Tang, W., Huang, S., Zhang, X., Zhou, F., Zhang, Y., Liu, B., 2023. Multiple instance learning framework with masked hard instance mining for whole slide image classification, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4078–4087

    2023

  37. [45]

    Tang, W., Zhou, F., Huang, S., Zhu, X., Zhang, Y., Liu, B., 2024. Feature re-embedding: Towards foundation model-level performance in computational pathology, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11343–11352

    2024

  38. [46]

    Attention is all you need

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. Advances in neural information processing systems 30

    2017

  39. [47]

    C-mil: Continuation multiple instance learning for weakly supervised object detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Wan, F., Liu, C., Ke, W., Ji, X., Jiao, J., Ye, Q., 2019. C-mil: Continuation multiple instance learning for weakly supervised object detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 2199–2208

    2019

  40. [48]

    Revisiting multiple instance neural networks

    Wang, X., Yan, Y., Tang, P., Bai, X., Liu, W., 2018. Revisiting multiple instance neural networks. Pattern recognition 74, 15–24

    2018

  41. [49]

    Pan-cancerhistopathologywsipre-trainingwithposition-aware masked autoencoder

    Wu,K.,Jiang,Z.,Tang,K.,Shi,J.,Xie,F.,Wang,W.,Wu,H.,Zheng,Y.,2024. Pan-cancerhistopathologywsipre-trainingwithposition-aware masked autoencoder. IEEE Transactions on Medical Imaging

    2024

  42. [50]

    Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y., 2022a. Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification, in: Proceedings of the IEEE/CVF conference on computer vision and pat...

  43. [51]

    Region-aware knowledge distillation for efficient image-to-image translation

    Zhang, L., Chen, X., Dong, R., Ma, K., 2022b. Region-aware knowledge distillation for efficient image-to-image translation. arXiv preprint arXiv:2205.12451

    arXiv

  44. [52]

    Vision mamba: Efficient visual representation learning with bidirectional state space model, in: Forty-first International Conference on Machine Learning

    Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X., . Vision mamba: Efficient visual representation learning with bidirectional state space model, in: Forty-first International Conference on Machine Learning. D.Cao et al.:Preprint submitted to Biomedical Signal Processi...

  45. [2021]

    Journal of big Data 8, 1–74

    Review of deep learning: concepts, cnn architectures, challenges, applications, future directions. Journal of big Data 8, 1–74

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.