Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Whole-slide cancer classification improves when entire clusters of non-diagnostic patches are pruned, not just individual patches.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-04 17:11 UTC pith:ROO6M5OX

load-bearing objection Cluster-level sparsity is a sensible idea, but the headline SOTA is picked on the test set and the theory is standard Lasso; salvageable, not publishable as is. the 3 major comments →

arxiv 2509.11034 v1 pith:ROO6M5OX submitted 2025-09-14 cs.CV

Cluster-Level Sparse Multi-Instance Learning for Whole-Slide Images

Yuedi Zhang , Zhixiang Xia , Guosheng Yin , Bin Liu This is my paper
classification cs.CV
keywords multi-instance learningwhole-slide imagescluster-level sparsityL1 regularizationattention mechanismcomputational pathologycompressed sensingweakly supervised classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that multi-instance learning on whole-slide images works better when pruning happens at the level of clusters, not individual patches. It groups every patch in a slide into K clusters defined by global centers, runs attention inside each cluster to form a prototype, then combines prototypes with learnable weights regularized by L1. A zero weight effectively discards an entire cluster, filtering out noisy tissue regions. The authors argue this needs O(s log K) slides to recover s relevant clusters, and report top accuracy, F1, and AUC on CAMELYON16 and TCGA-NSCLC.

Core claim

The central claim is that sparse selection of cluster-level prototypes—not instance-level attention alone—is what makes weakly supervised whole-slide classification robust. csMIL creates K global cluster centers from all training patches, assigns each bag's patches to those centers, computes attention-weighted prototypes per cluster, and sums the prototypes with a K-dimensional weight vector β regularized by ℓ1. In the learned model, clusters whose β components shrink to zero are discarded, so each slide is classified from a small set of diagnostically relevant tissue patterns. The paper reports that this beats attention-based and instance-selection baselines on CAMELYON16 (0.905 accuracy, 0

What carries the argument

The load-bearing object is the cluster-level sparse aggregation: K global cluster centers, within-cluster attention (ABMIL-style) that produces one prototype per cluster, and a shared sparse weight vector β with ℓ1 regularization. The zero or near-zero entries of β are the selection mechanism: they mark clusters of patches the model discards. The theory maps this to Lasso and compressed sensing, with the bag-level cluster prototypes as the design matrix and a restricted eigenvalue or incoherence condition as the bridge.

Load-bearing premise

The method assumes that diagnostically relevant and irrelevant tissue patches sort into distinct clusters, and that the cluster-aggregated features satisfy the restricted eigenvalue and incoherence conditions; if either fails, the sparsity-based pruning and the O(s log K) guarantee collapse.

What would settle it

Run csMIL on a synthetic MIL dataset where positive instances are deliberately spread across every cluster; if recovering the signal requires more than O(s log K) bags or forces nonzero weights on all clusters, the separation assumption is falsified. On real slides, a simpler check: take a trained model, remove all patches from zero-weight clusters on held-out slides, and see whether predictions stay the same—if they shift, the sparsity mask is not truly selecting irrelevant tissue.

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

If this is right

  • If the central claim holds, whole-slide classifiers can drop entire non-diagnostic tissue classes without retraining, because β learns to zero them out.
  • The O(s log K) recovery statement means the number of slides needed scales only logarithmically with the number of clusters, so finer-grained clustering does not force proportionally more data.
  • Cluster-level sparsity maps give pathologists a direct visual: retained clusters are diagnostic regions, zero-weight clusters are noise, which is the paper's interpretability story.
  • Because K is much smaller than the number of instances, the per-step cost avoids the quadratic instance-pair cost of transformer MIL and the instance-level scoring cost of selection-based MIL.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: If the separation premise holds, the same cluster-pruning idea could transfer to other weakly labeled bag problems where relevant and irrelevant patterns are spatially coherent—a testable extension the paper does not run.
  • Editorial inference: The bound's dependence on learned features is the key gap; since the prototypes are data-dependent, the restricted eigenvalue and incoherence constants are not guaranteed, so a natural extension is to verify the eigengap on real feature embeddings.
  • Editorial inference: The paper's K-sensitivity results suggest a practical heuristic: choose K by the stability of β's support rather than final accuracy, since over-segmentation diluted features at K = 5 and K = 10.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes csMIL, a multi-instance learning framework for whole-slide images that combines global K-means clustering over all instances, local assignment of each bag's instances to the K global centers, within-cluster attention pooling into K prototypes, and a bag-level aggregation z_i = sum_k beta_k z_i^k with an l1 penalty on beta. The authors argue that cluster-level sparsity discards irrelevant instance cohorts, improves interpretability, and reduces computational cost. They state a sparse-recovery guarantee (Theorem 1: O(s log K) bags suffice to recover s relevant clusters) and report state-of-the-art accuracy/F1/AUC on CAMELYON16 (0.905/0.876/0.951) and TCGA-NSCLC (0.854/0.855/0.933), together with qualitative pathology-informed visualizations of cluster weights.

Significance. If the empirical claim were supported, csMIL would be a simple and attractive addition to attention-based MIL: it repurposes standard K-means and l1 regularization into an interpretable cluster-selection mechanism, and the qualitative maps in Section 5.5 are a genuinely useful clinical-interpretability direction. The paper, however, does not currently provide a sound empirical foundation for the SOTA claim. The lack of a validation split and the explicit test-set-driven sweeps over K and gamma (Tables 3–4, Figure 5) bias the headline numbers upward. The theoretical section is also not a derivation for the actual model: the proof assumes squared-error linear regression and assumes, rather than establishes, restricted eigenvalue and incoherence conditions on learned cluster features; the appendix itself concedes that nonlinear dependencies and poor clustering invalidate the bound. The central insight is potentially publishable, but the current evidence is not sufficient.

major comments (3)
  1. [§5.1, §5.4, §5.6; Tables 1–4; Fig. 5] The reported SOTA is obtained by test-set-driven hyperparameter selection. The protocol in §5.1 has only training and test folds (4:1), with no validation split. §5.4 sweeps K in {2,3,5,10} and reports test-fold Acc/F1/AUC (Tables 3–4); §5.6 sweeps 15 values of gamma and reports test accuracy (Figure 5). The final Tables 1–2 then use K=3 and the best gamma found in these sweeps. Selecting hyperparameters on the same test folds used for comparison makes the headline numbers optimistically biased and makes comparisons to baselines unfair. The paper must either use a proper validation/nested-CV procedure or clearly report the selection mechanism and per-fold variance; without this, the SOTA claim is unsupported.
  2. [§4.3 and Appendix (Proof of Theorem 1)] Theorem 1 as stated does not follow from the actual objective. The appendix proof analyzes the Lasso problem min_beta (1/(2M)) sum_i (Y_i - beta^T z_i)^2 + gamma ||beta||_1, while the method's loss in Eq. (4) is sum_i CE(g(z_i), Y_i) + gamma ||beta||_1, with g mapping z_i to a probability simplex. The theorem is therefore a generic Lasso sample-complexity statement, not a recovery guarantee for csMIL as implemented. Moreover, the restricted eigenvalue and incoherence conditions are assumed on learned cluster-aggregated features rather than verified; the appendix itself states that nonlinear dependencies and poor clustering invalidate the bound. The authors should either prove the bound for the actual cross-entropy objective under explicit, checkable assumptions or substantially reframe the theoretical contribution.
  3. [§3.2 and §4.3] The method and the theory both rest on the assumption that diagnostically relevant and irrelevant instances reside in different clusters. The only support offered is the qualitative bar plot in Figure 1, with no cluster-purity, silhouette, or other cluster-quality metric, and no ablation in which this assumption is artificially violated. Since the pruning logic and the O(s log K) guarantee collapse if clusters mix relevant and irrelevant patterns or if learned features are collinear, the paper should quantify cluster quality on the actual learned features and test robustness to imperfect clustering.
minor comments (5)
  1. [Eq. (3) and §3.3] The notation in Eq. (3) is hard to parse (h_n^T inside tanh), and the summation index in z_i^k = sum_{i=1}^{C_i^k} alpha_i^k h_i reuses i for both the bag index and the instance index. Please reindex instances (e.g., j) and clarify the dimensions.
  2. [§5.4] The opening sentence says the K-sweep is 'shown in Table 1 and 2 respectively,' but the actual K-sweep results are in Tables 3 and 4. Please correct the cross-reference.
  3. [Figure 5 and §5.6] The caption is missing the gamma symbol ('Model Accuracy vs. [gamma]'), and the text says 'accuracy increases from 0.750 to 0.900' when it presumably means accuracy rises over the gamma range. Please clarify the axis description and the direction of the claim.
  4. [References and dataset description] The TCGA-NSCLC dataset description (§5.1) does not match reference [54], which is a breast-cancer TCGA paper; a lung-cancer TCGA reference should be cited. Also, the abstract contains 'CAMEL YON16' (spacing typo).
  5. [Appendix] The appendix contains a placeholder sentence 'Appendix two text goes here' followed by the acknowledgment line. This should be removed before submission.

Circularity Check

1 steps flagged

Reported SOTA is selected on the test set: K and γ are tuned on the same folds used for evaluation, so Tables 1-2 overstate csMIL's advantage; the theoretical bound is standard Lasso with explicitly assumed conditions.

specific steps
  1. fitted input called prediction [Section 5.4 (Tables 3-4), Section 5.6 (Fig. 5), feeding Section 5.3 Tables 1-2; protocol Section 5.1]
    "The number of clusters, K, is a hyperparameter that will be selected based on the final classification performance of the bags. ... First, the optimal number of clusters, denoted as K = 3, achieves the highest performance across all evaluated metrics, specifically an accuracy of 0.905, an F1 score of 0.876, and an AUC of 0.951. ... Parameter γ selection via network search."

    Section 5.1 partitions each dataset into only training and test sets (4:1) with no validation set. Section 3.3 says K is selected based on final classification performance; Section 5.4 sweeps K and reports the best row, and Section 5.6 sweeps γ on the same data (Figure 5). Tables 1-2 then report the K=3 and best-γ result as 'Our method'. Hence the headline Acc/F1/AUC are the maximum of a hyperparameter sweep evaluated on the same test folds used for comparison, not independent out-of-sample predictions. The reported numbers are fitted to the test labels by construction, so the SOTA claim is statistically forced.

full rationale

One concrete circularity is present in the empirical headline claim. The protocol (Sec. 5.1) has only train/test splits, and Sec. 3.3 states K is chosen by final classification performance; Sec. 5.4 then sweeps K and identifies K=3 as the best on the same test folds, and Sec. 5.6 sweeps γ on the same data. Tables 1-2 report the K=3 row and the best-γ result as 'Our method,' so the reported SOTA is the maximum of a hyperparameter search on the evaluation data. That is a fitted-input-called-prediction pattern: the prediction (test accuracy) is, by construction, the outcome used to select the hyperparameters, making the headline numbers optimistically biased and not independent predictions. The theoretical O(s log K) claim is different: it is a direct restatement of standard Lasso sample-complexity bounds, with the paper's own Appendix explicitly limiting the claim to the case where the learned cluster features satisfy RE/incoherence and a linear label model. This is an assumption, not a circular derivation, so I do not count it as a separate circular step. The self-citations ([14], [41]) are background and not load-bearing. Score 6: the empirical central claim reduces, in part, to test-set-driven selection; the architecture and the conditional theory retain independent content.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The method's central claim rests on unverified statistical and clustering-structure axioms, plus two hyperparameters (K and γ) tuned on the test set. These are the load-bearing assumptions the reader must accept for the reported SOTA and recovery bound to hold.

free parameters (2)
  • K (number of clusters) = 3 (selected as best on both datasets)
    Tuned via test-set performance in Section 5.4; K=2,5,10 give worse numbers.
  • γ (ℓ1 regularization weight) = Not stated explicitly; selected from grid {1e-4, ..., 0.1} by accuracy on the test set
    Section 5.6 and Figure 5 search 15 values and report the best accuracy, with no held-out validation split.
axioms (4)
  • domain assumption Positive and negative instances reside in distinct clusters.
    Section 3.2 uses this hypothesis to justify the entire cluster-level sparsity design; no evidence is provided that the K-means feature space separates diagnostic from non-diagnostic patches.
  • ad hoc to paper Learned cluster-aggregated features satisfy the Restricted Eigenvalue and incoherence conditions.
    The Appendix proof relies on these unverified conditions to apply Lasso recovery bounds to the network's learned z_i features.
  • ad hoc to paper Bag labels are linearly predicted by cluster prototypes: Y_i = β^T z_i + ε_i.
    The Appendix assumes a linear model with sub-Gaussian noise, but the actual training uses cross-entropy on a nonlinear attention network, so the theory does not match the method.
  • domain assumption Global K-means centers transfer to each bag for local cluster assignment.
    Section 3.3 assumes that centers computed across all training bags are meaningful anchors for every individual slide, an assumption that can fail for rare tissue types.

pith-pipeline@v1.3.0-alltime-deepseek · 17150 in / 11620 out tokens · 121008 ms · 2026-08-04T17:11:11.798095+00:00 · methodology

0 comments
read the original abstract

Multi-Instance Learning (MIL) is pivotal for analyzing complex, weakly labeled datasets, such as whole-slide images (WSIs) in computational pathology, where bags comprise unordered collections of instances with sparse diagnostic relevance. Traditional MIL approaches, including early statistical methods and recent attention-based frameworks, struggle with instance redundancy and lack explicit mechanisms for discarding non-informative instances, limiting their robustness and interpretability. We propose Cluster-level Sparse MIL (csMIL), a novel framework that integrates global-local instance clustering, within-cluster attention, and cluster-level sparsity induction to address these challenges. Our csMIL first performs global clustering across all bags to establish $K$ cluster centers, followed by local clustering within each bag to assign cluster labels. Attention scores are computed within each cluster, and sparse regularization is applied to cluster weights, enabling the selective retention of diagnostically relevant clusters while discarding irrelevant ones. This approach enhances robustness to noisy instances, improves interpretability by identifying critical regions, and reduces computational complexity. Theoretical analysis demonstrates that csMIL requires $O(s log K)$ bags to recover $s$ relevant clusters, aligning with compressed sensing principles. Empirically, csMIL achieves state-of-the-art performance on two public histopathology benchmarks (CAMELYON16, TCGA-NSCLC).

Figures

Figures reproduced from arXiv: 2509.11034 by Bin Liu, Guosheng Yin, Yuedi Zhang, Zhixiang Xia.

Figure 1
Figure 1. Figure 1: Under the framework of ABMIL, all samples are sliced to obtain instances, and these instances are clustered into five groups. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparisons with the original ABMIL model under the setting of [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The clustering visualization of one patient’s whole-slide image. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of patches under different clustering settings [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Parameter γ selection via network search. 6 CONCLUSION Multi-instance learning (MIL) has gained significant traction re￾cently due to its ability to handle weakly labeled data, making it particularly relevant in domains where precise instance-level annotations are costly or impractical. The csMIL framework introduced in this study represents a significant advancement in addressing the challenges of MIL for… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 4 linked inside Pith

  1. [1]

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

    M. Y . Lu, D. F. Williamson, T. Y . Chen, R. J. Chen, M. Barbieri, and F. Mahmood, “Data-efficient and weakly supervised computational pathology on whole-slide images,” Nature Biomedical Engineering , vol. 5, no. 6, pp. 555–570, 2021

    2021

  2. [2]

    Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential,

    M. Gadermayr and M. Tschuchnig, “Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential,” Computerized Medical Imaging and Graphics , vol. 112, p. 102337, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0895611124000144

    2024

  3. [3]

    Pseudo-bag mixup augmentation for multiple instance learning-based whole slide image classification,

    P. Liu, L. Ji, X. Zhang, and F. Ye, “Pseudo-bag mixup augmentation for multiple instance learning-based whole slide image classification,” IEEE Transactions on Medical Imaging, vol. 43, no. 5, pp. 1841–1852, 2024

    2024

  4. [4]

    Advmil: Adversarial multiple instance learning for the survival analysis on whole-slide images,

    P. Liu, L. Ji, F. Ye, and B. Fu, “Advmil: Adversarial multiple instance learning for the survival analysis on whole-slide images,” Medical Image Analysis, vol. 91, p. 103020, 2024

    2024

  5. [5]

    A framework for multiple-instance learning,

    O. Maron and T. Lozano-P ´erez, “A framework for multiple-instance learning,” Advances in neural information processing systems , vol. 10, 1997

    1997

  6. [6]

    Support vector ma- chines for multiple-instance learning,

    S. Andrews, I. Tsochantaridis, and T. Hofmann, “Support vector ma- chines for multiple-instance learning,” Advances in neural information processing systems, vol. 15, 2002

    2002

  7. [7]

    Multiple instance learning convolutional neural networks for object recognition,

    M. Sun, T. X. Han, M.-C. Liu, and A. Khodayari-Rostamabad, “Multiple instance learning convolutional neural networks for object recognition,” in 2016 23rd International Conference on Pattern Recognition (ICPR) . IEEE, 2016, pp. 3270–3275

    2016

  8. [8]

    Attention-based deep multiple instance learning,

    M. Ilse, J. Tomczak, and M. Welling, “Attention-based deep multiple instance learning,” in Proceedings of the 35th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, J. Dy and A. Krause, Eds., vol. 80. PMLR, 10–15 Jul 2018, pp. 2127– 2136

    2018

  9. [9]

    Loss-based attention for deep multiple instance learning,

    X. Shi, F. Xing, Y . Xie, Z. Zhang, L. Cui, and L. Yang, “Loss-based attention for deep multiple instance learning,” inProceedings of the AAAI conference on artificial intelligence , vol. 34, no. 04, 2020, pp. 5742– 5749

    2020

  10. [10]

    Multi- instance partial-label learning with margin adjustment,

    W. Tang, Y .-F. Yang, Z. Wang, W. Zhang, and M.-L. Zhang, “Multi- instance partial-label learning with margin adjustment,” Advances in Neural Information Processing Systems , vol. 37, pp. 26 331–26 354, 2024

    2024

  11. [11]

    Miles: Multiple-instance learning via embedded instance selection,

    Y . Chen, J. Bi, and J. Wang, “Miles: Multiple-instance learning via embedded instance selection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 12, pp. 1931–1947, 2006

    1931

  12. [12]

    Adaptis: Adaptive instance selection network,

    K. Sofiiuk, O. Barinova, and A. Konushin, “Adaptis: Adaptive instance selection network,” in Proceedings of the IEEE/CVF international con- ference on computer vision, 2019, pp. 7355–7363

    2019

  13. [13]

    Deep multiple instance selection,

    X.-C. Li, D.-C. Zhan, J.-Q. Yang, and Y . Shi, “Deep multiple instance selection,” Science China Information Sciences, vol. 64, pp. 1–15, 2021

    2021

  14. [14]

    Sparse multiple instance learning for elderly people balance ability,

    Y . Zhang, D. Su, X. Zhao, and Y . Tian, “Sparse multiple instance learning for elderly people balance ability,” Procedia Computer Science, vol. 199, pp. 621–628, 2022

    2022

  15. [15]

    Primary mucinous ovarian tumors vs. ovarian metastases from gastrointestinal tract, pancreas and biliary tree: a review of current problematics,

    P. Dundr, N. Singh, B. No ˇziˇckov´a, K. N ˇemejcov´a, M. B ´art˚u, and I. Struˇzinsk´a, “Primary mucinous ovarian tumors vs. ovarian metastases from gastrointestinal tract, pancreas and biliary tree: a review of current problematics,” Diagnostic pathology, vol. 16, pp. 1–17, 2021

    2021

  16. [16]

    Pathology of primary and metastatic mucinous ovarian neoplasms,

    S. L. S. Leen and N. Singh, “Pathology of primary and metastatic mucinous ovarian neoplasms,” Journal of Clinical Pathology , vol. 65, no. 7, pp. 591–595, 2012

    2012

  17. [17]

    Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,

    H. Zhang, Y . Meng, Y . Zhao, Y . Qiao, X. Yang, S. E. Coupland, and Y . Zheng, “Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,” in Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 18 802–18 812

    2022

  18. [18]

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

    G. Campanella, M. G. Hanna, L. Geneslaw, A. Miraflor, V . Werneck Krauss Silva, K. J. Busam, E. Brogi, V . E. Reuter, D. S. Klimstra, and T. J. Fuchs, “Clinical-grade computational pathology using weakly supervised deep learning on whole slide images,” Nature Medicine, vol. 25, no. 8, pp. 1301–1309, 2019

    2019

  19. [19]

    Solving the multi- ple instance problem with axis-parallel rectangles,

    T. G. Dietterich, R. H. Lathrop, and T. Lozano-P ´erez, “Solving the multi- ple instance problem with axis-parallel rectangles,”Artificial intelligence, vol. 89, no. 1-2, pp. 31–71, 1997

    1997

  20. [20]

    Multi instance neural networks,

    J. Ramon and L. De Raedt, “Multi instance neural networks,” in Pro- ceedings of the ICML-2000 workshop on attribute-value and relational learning, 2000, pp. 53–60

    2000

  21. [21]

    Key instance detection in multi-instance learning,

    G. Liu, J. Wu, and Z.-H. Zhou, “Key instance detection in multi-instance learning,” in Asian conference on machine learning . PMLR, 2012, pp. 253–268

    2012

  22. [22]

    Multiple instance learning: A survey of problem characteristics and applications,

    M.-A. Carbonneau, V . Cheplygina, E. Granger, and G. Gagnon, “Multiple instance learning: A survey of problem characteristics and applications,” Pattern recognition, vol. 77, pp. 329–353, 2018

    2018

  23. [23]

    From group to individual labels using deep features,

    D. Kotzias, M. Denil, N. De Freitas, and P. Smyth, “From group to individual labels using deep features,” in Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, 2015, pp. 597–606

    2015

  24. [24]

    Support vector ma- chines for multiple-instance learning,

    S. Andrews, I. Tsochantaridis, and T. Hofmann, “Support vector ma- chines for multiple-instance learning,” in Proceedings of the 16th In- ternational Conference on Neural Information Processing Systems , ser. NIPS’02. Cambridge, MA, USA: MIT Press, 2002, p. 577–584

    2002

  25. [25]

    Em-dd: An improved multiple-instance learning technique,

    Q. Zhang and S. Goldman, “Em-dd: An improved multiple-instance learning technique,” Advances in neural information processing systems, vol. 14, 2001

    2001

  26. [26]

    Deep sets,

    M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdinov, and A. J. Smola, “Deep sets,” Advances in neural information processing systems, vol. 30, 2017

    2017

  27. [27]

    Deep multiple instance learning for image classification and auto-annotation,

    J. Wu, Y . Yu, C. Huang, and K. Yu, “Deep multiple instance learning for image classification and auto-annotation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 3460– 3469

    2015

  28. [28]

    Trans- MIL: Transformer based correlated multiple instance learning for whole slide image classification,

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

    2021

  29. [29]

    Multi-view multi-instance learning based on joint sparse representation and multi-view dictionary learning,

    B. Li, C. Yuan, W. Xiong, W. Hu, H. Peng, X. Ding, and S. Maybank, “Multi-view multi-instance learning based on joint sparse representation and multi-view dictionary learning,” IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, vol. 39, no. 12, pp. 2554–2560, 2017

    2017

  30. [30]

    Weakly supervised object lo- calization with multi-fold multiple instance learning,

    R. G. Cinbis, J. Verbeek, and C. Schmid, “Weakly supervised object lo- calization with multi-fold multiple instance learning,” IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 1, pp. 189–203, 2016

    2016

  31. [31]

    Multi-layer multi-instance learning for video concept detection,

    Z. Gu, T. Mei, X.-S. Hua, J. Tang, and X. Wu, “Multi-layer multi-instance learning for video concept detection,” IEEE Transactions on Multimedia, vol. 10, no. 8, pp. 1605–1616, 2008

    2008

  32. [32]

    Video anomaly detection with ntcn-ml: A novel tcn for multi-instance learning,

    W. Shao, R. Xiao, P. Rajapaksha, M. Wang, N. Crespi, Z. Luo, and R. Minerva, “Video anomaly detection with ntcn-ml: A novel tcn for multi-instance learning,” Pattern Recognition, vol. 143, p. 109765, 2023

    2023

  33. [33]

    Text representation and classification based on multi-instance learning,

    W. He and Y . Wang, “Text representation and classification based on multi-instance learning,” in 2009 International Conference on Manage- ment Science and Engineering. IEEE, 2009, pp. 34–39

    2009

  34. [34]

    Multiple instance learning for classification of dementia in brain mri,

    T. Tong, R. Wolz, Q. Gao, R. Guerrero, J. V . Hajnal, D. Rueckert, A. D. N. Initiative et al. , “Multiple instance learning for classification of dementia in brain mri,” Medical image analysis , vol. 18, no. 5, pp. 808–818, 2014

    2014

  35. [35]

    Classification and disease localization in histopathology using only global labels: A weakly-supervised approach,

    P. Courtiol, E. W. Tramel, M. Sanselme, and G. Wainrib, “Classification and disease localization in histopathology using only global labels: A weakly-supervised approach,” arXiv preprint arXiv:1802.02212, 2018

    Pith/arXiv arXiv 2018

  36. [36]

    Whole slide images based cancer survival prediction using attention guided deep JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 12 multiple instance learning networks,

    J. Yao, X. Zhu, J. Jonnagaddala, N. Hawkins, and J. Huang, “Whole slide images based cancer survival prediction using attention guided deep JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 12 multiple instance learning networks,” Medical image analysis, vol. 65, p. 101789, 2020

    2015

  37. [37]

    Dual-stream multiple instance learn- ing network for whole slide image classification with self-supervised contrastive learning,

    B. Li, Y . Li, and K. W. Eliceiri, “Dual-stream multiple instance learn- ing network for whole slide image classification with self-supervised contrastive learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 318–14 328

    2021

  38. [38]

    Murcl: Multi-instance reinforcement contrastive learning for whole slide image classification,

    Z. Zhu, L. Yu, W. Wu, R. Yu, D. Zhang, and L. Wang, “Murcl: Multi-instance reinforcement contrastive learning for whole slide image classification,” IEEE Transactions on Medical Imaging , vol. 42, no. 5, pp. 1337–1348, 2022

    2022

  39. [39]

    Advances in multiple instance learning for whole slide image analysis: Techniques, challenges, and future directions,

    J. Wang, Y . Mao, N. Guan, and C. J. Xue, “Advances in multiple instance learning for whole slide image analysis: Techniques, challenges, and future directions,” arXiv preprint arXiv:2408.09476, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Rethinking attention-based multiple instance learning for whole-slide pathological image classification: An instance attribute viewpoint,

    L. Cai, S. Huang, Y . Zhang, J. Lu, and Y . Zhang, “Rethinking attention-based multiple instance learning for whole-slide pathological image classification: An instance attribute viewpoint,” arXiv preprint arXiv:2404.00351, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    Aem: At- tention entropy maximization for multiple instance learning based whole slide image classification,

    Y . Zhang, Z. Shui, Y . Sun, H. Li, J. Li, C. Zhu, and L. Yang, “Aem: At- tention entropy maximization for multiple instance learning based whole slide image classification,” arXiv preprint arXiv:2406.15303, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Rethinking multiple instance learning for whole slide image classification: A good instance classifier is all you need,

    L. Qu, Y . Ma, X. Luo, Q. Guo, M. Wang, and Z. Song, “Rethinking multiple instance learning for whole slide image classification: A good instance classifier is all you need,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 10, pp. 9732–9744, 2024

    2024

  43. [43]

    Attention is not what you need: Revisiting multi-instance learning for whole slide image classification,

    X. Liu, W. Zhang, and M.-L. Zhang, “Attention is not what you need: Revisiting multi-instance learning for whole slide image classification,” arXiv preprint arXiv:2408.09449, 2024

    arXiv 2024

  44. [44]

    Hmil: Hierarchical multi- instance learning for fine-grained whole slide image classification,

    C. Jin, L. Luo, H. Lin, J. Hou, and H. Chen, “Hmil: Hierarchical multi- instance learning for fine-grained whole slide image classification,”IEEE Transactions on Medical Imaging, vol. 44, no. 4, pp. 1796–1808, 2025

    2025

  45. [45]

    Camil: channel attention-based multiple instance learning for whole slide image classification,

    J. Mao, J. Xu, X. Tang, Y . Liu, H. Zhao, G. Tian, and J. Yang, “Camil: channel attention-based multiple instance learning for whole slide image classification,” Bioinformatics, vol. 41, no. 2, p. btaf024, 2025

    2025

  46. [46]

    Targeting tumor heterogeneity: multiplex- detection-based multiple instance learning for whole slide image clas- sification,

    Z. Wang, Y . Bi, T. Pan, X. Wang, C. Bain, R. Bassed, S. Imoto, J. Yao, R. J. Daly, and J. Song, “Targeting tumor heterogeneity: multiplex- detection-based multiple instance learning for whole slide image clas- sification,” Bioinformatics, vol. 39, no. 3, p. btad114, 2023

    2023

  47. [47]

    Camil: Causal multiple instance learning for whole slide image classification,

    K. Chen, S. Sun, and J. Zhao, “Camil: Causal multiple instance learning for whole slide image classification,” in Proceedings of the AAAI Con- ference on Artificial Intelligence, vol. 38, no. 2, 2024, pp. 1120–1128

    2024

  48. [48]

    Multiple instance learning with random sampling for whole slide image classification,

    H. Keshvarikhojasteh, J. P. Pluim, and M. Veta, “Multiple instance learning with random sampling for whole slide image classification,” in Medical Imaging 2024: Digital and Computational Pathology , vol. 12933. SPIE, 2024, pp. 372–376

    2024

  49. [49]

    Learning with ℓ1-graph for image analysis,

    B. Cheng, J. Yang, S. Yan, Y . Fu, and T. S. Huang, “Learning with ℓ1-graph for image analysis,” IEEE Transactions on Image Processing , vol. 19, no. 4, pp. 858–866, 2010

    2010

  50. [50]

    Multi-instance learning by treating instances as non-i.i.d. samples,

    Z.-H. Zhou, Y .-Y . Sun, and Y .-F. Li, “Multi-instance learning by treating instances as non-i.i.d. samples,” in Proceedings of the 26th Annual Inter- national Conference on Machine Learning , ser. ICML ’09. New York, NY , USA: Association for Computing Machinery, 2009, p. 1249–1256

    2009

  51. [51]

    Adaptive prototype learning and allocation for few-shot segmentation,

    G. Li, V . Jampani, L. Sevilla-Lara, D. Sun, J. Kim, and J. Kim, “Adaptive prototype learning and allocation for few-shot segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8334–8343

    2021

  52. [52]

    Enhancing sparsity by reweighted l 1 minimization,

    E. J. Candes, M. B. Wakin, and S. P. Boyd, “Enhancing sparsity by reweighted l 1 minimization,” Journal of Fourier analysis and applica- tions, vol. 14, pp. 877–905, 2008

    2008

  53. [53]

    Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer,

    B. E. Bejnordi, M. Veta, P. J. Van Diest, B. Van Ginneken, N. Karsse- meijer, G. Litjens, J. A. Van Der Laak, M. Hermsen, Q. F. Manson, M. Balkenhol et al., “Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer,” Jama, vol. 318, no. 22, pp. 2199–2210, 2017

    2017

  54. [54]

    Comprehensive molecular portraits of human breast tumours,

    C. G. A. Network, “Comprehensive molecular portraits of human breast tumours,” Nature, vol. 490, no. 7418, pp. 61–70, 2012

    2012

  55. [55]

    Set trans- former: A framework for attention-based permutation-invariant neural networks,

    J. Lee, Y . Lee, J. Kim, A. Kosiorek, S. Choi, and Y . W. Teh, “Set trans- former: A framework for attention-based permutation-invariant neural networks,” in International conference on machine learning . PMLR, 2019, pp. 3744–3753

    2019

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.