Pith. sign in

REVIEW 3 major objections 4 minor 47 references

Hypergraph Mamba for Efficient Whole Slide Image Understanding

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read WSI-HGMamba combines hypergraph message passing with a bidirectional state space model to classify whole-slide images at Transformer-level accuracy with up to 7x lower FLOPs.

desk verdict Plausible hypergraph-Mamba MIL design with promising accuracy, but the linear-efficiency claim ignores the quadratic cost of similarity-based hyperedge construction. read the letter →

arxiv 2505.17457 v2 pith:WG4IRA47 submitted 2025-05-23 cs.CV cs.AI

classification cs.CVcs.AI
keywords wholeslideimageclassificationmultipleinstancelearninghypergraphneuralnetworkstatespacemodelMambacomputationalpathologyhistopathology
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

The paper sets out to show that whole-slide image classification does not have to choose between relational modeling and efficient sequence modeling. It introduces WSI-HGMamba, which builds a hypergraph over tissue tiles from both spatial adjacency and top-K similarity hyperedges, flattens that hypergraph into ordered node sequences, and processes the sequences with a bidirectional state space model. The reported result is that this hybrid matches or exceeds Transformer and Graph Transformer baselines on TCGA-ESCA, TCGA-NSCLC, TCGA-RCC, and an in-house prostate dataset while using up to 7x fewer FLOPs. If true, this gives pathology AI a backbone that can handle tens of thousands of tiles per slide without the quadratic cost of attention.

What carries the argument

The load-bearing component is the HGMamba block. It takes node features $X$ and a unified hypergraph incidence matrix $H = [H_{\text{rule}}, H_{\text{sim}}]$, where $H_{\text{rule}}$ connects spatially adjacent tiles and $H_{\text{sim}}$ connects each tile to its top-$K$ most cosine-similar tiles. The block applies hypergraph convolution $X^{(l)} = \sigma\big(D_v^{-1/2} H W_e D_e^{-1} H^\top D_v^{-1/2} X W^{(l-1)}\big)$, then converts the hypergraph into $M$ node sequences using hypergraph depth-first search and acyclic random walk, then processes each sequence with a bidirectional state space model that includes 1D convolution, SSM, normalization, and residual addition. An aggregator averages each node's representations across all sequences, returning the features to hypergraph form for the next block.

What would settle it

Measure the full per-slide pipeline, including the top-K cosine-similarity hyperedge construction, against the tile count N; if end-to-end FLOPs or wall-clock time grows quadratically rather than linearly, the central scalability claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that high-order tile relationships and long-range contextual dependencies can be captured in one linear-time architecture. On four slide-level classification benchmarks, the model reports AUCs of 98.7%, 98.4%, 99.3%, and 98.8%, above the compared Transformer and Graph Transformer baselines, while using 2.2 GFLOPs and 1.0 GB of memory; one graph-transformer counterpart reports 12.5 GFLOPs and 2.5 GB. The authors attribute the gain to the HGMamba block, where hypergraph message passing preserves spatial and semantic structure, and bidirectional state space modeling carries context across long sequences.

Load-bearing premise

The linear-time efficiency claim assumes that building the similarity-based hyperedges is cheap, but as described, finding the top-K most cosine-similar tiles for each tile is quadratic in the number of tiles, and no approximate neighbor search is specified.

Editorial extensions

If this is right

  • Slide-level classification on a single commodity GPU becomes practical for slides with tens of thousands of tiles, since the SSM path is linear in sequence length and measured FLOPs stay near GNN levels.
  • Bidirectional context is consistently worth more than a unidirectional SSM, and spatial plus semantic hyperedges are complementary, so both design choices should carry over to other WSI models.
  • Structured traversal matters: combining hypergraph depth-first search with acyclic random walks beats either scan alone and beats random ordering, meaning how a hypergraph is flattened into sequences is itself part of the modeling.
  • The block is more scalable in depth than plain GCN because the SSM path mitigates over-smoothing, so deeper hypergraph-SSM stacks remain effective.

Reading between the lines

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

  • The paper's linear-time claim appears to count the SSM and message-passing stages but not the construction of similarity hyperedges; finding top-$K$ cosine neighbors for each of $N$ tiles is $O(N^2 d)$ as written, so end-to-end linear scaling would require an approximate nearest-neighbor scheme or learned hyperedge assignment.
  • The flatten-and-scan recipe is not pathology-specific: any graph or hypergraph with node features could be scanned into sequences and processed by an SSM, so the block may transfer to other gigapixel-scale prediction tasks such as spatial transcriptomics or 3D pathology.
  • Because the aggregator averages each node's tokens over $M$ traversals, performance likely depends on how well the scans cover the hypergraph; a testable extension is choosing traversal roots or sequence counts adaptively per slide.
  • A direct comparison that includes end-to-end preprocessing time, rather than FLOPs of the network alone, would clarify whether the efficiency advantage survives in practice.
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

3 major / 4 minor

Summary. The manuscript proposes WSI-HGMamba, a whole-slide image classification framework that combines hypergraph neural network message passing with bidirectional state space sequence modeling. Tiles are embedded with a pre-trained ResNet50, a hypergraph is built from rule-based spatial edges and similarity-based top-K hyperedges, and a stack of HGMamba blocks alternates hypergraph convolution, hypergraph traversal (H-DFS and H-ARW) into M sequences, and Bi-SSM processing, followed by ABMIL classification. Experiments on TCGA-ESCA, TCGA-NSCLC, TCGA-RCC and a private Prost dataset report state-of-the-art AUC/ACC/F1 values with lower FLOPs than Transformer and Graph Transformer baselines.

Significance. The architectural idea of converting a hypergraph into traversal sequences and processing them with a bidirectional SSM is interesting and potentially useful for gigapixel pathology, and the paper includes a broad baseline comparison and several ablations. However, the central efficiency claim is not yet established end-to-end: the similarity-based hyperedge construction in Eqs. (4)-(5) is O(N^2 d) and is excluded from the reported FLOPs and from Fig. 3(a,b), so the linear-complexity and 7x-FLOPs-reduction conclusions are premature. In addition, key hyperparameters (K=3, number of layers=2) are selected using ablations on the same Prost data later used for reporting, and the quantitative results are presented on a single split without error bars. These issues are fixable, but they currently place the quantitative claims on weaker ground than the paper's conclusion suggests.

major comments (3)
  1. [WSI Hypergraph Construction (Eqs. (4)-(5))] The similarity-based adjacency construction requires finding the top-K most cosine-similar tiles among all N tiles for each tile. Without an approximate-nearest-neighbor scheme or a complexity analysis, this step costs O(N^2 d) multiply-adds and O(N^2 log K) comparisons. For a WSI with N=10,000 tiles and d=2048 features, the pairwise dot products alone are roughly 4e11 FLOPs, about 190x the reported 2.2 GFLOPs in Table 1. Since Fig. 3(a,b) and the FLOPs column count only the HGMamba network and not the construction of H_sim, the conclusion that the framework maintains linear computational complexity is not supported for the end-to-end pipeline. Please add the construction cost, use an efficient ANN implementation, or restrict the efficiency claims to the network forward pass.
  2. [Ablation on Top-K Neighbors / Ablation on HGMamba Layer] The main reported configuration uses K=3 and two HGMamba layers, but these values are chosen based on Fig. 3(c) and Fig. 3(f), which evaluate performance on the same Prost dataset used for final reporting. There is no indication that a held-out validation split was used for this selection, and the results in Tables 1 and 2 are reported on a single split without standard deviations. The accuracy advantages over strong baselines may therefore be inflated by selection bias and split luck. Please adopt a validation-based hyperparameter selection protocol, report test-set numbers with multiple random splits and error bars, and state the exact split protocol for the private dataset.
  3. [Abstract / Conclusion / Table 1] The abstract and conclusion state that WSI-HGMamba achieves up to 7x reduction in FLOPs relative to Transformer and Graph Transformer counterparts. In Table 1, the heaviest Graph Transformer baseline is IGT at 12.5 GFLOPs, which is only 12.5/2.2 ≈ 5.7x the FLOPs of WSI-HGMamba; the intermediate WSI-GMamba reaches 12.5/1.8 ≈ 6.9x. If the 7x figure is intended for WSI-GMamba or is a rounded value, the text should say so; otherwise the claim is not supported by the reported numbers.
minor comments (4)
  1. [HGMamba Block, Eq. (7)] The update equation applies the hypergraph convolution to X rather than to X^{(l-1)}, and the text has a typo 'parameter matrixs'. Please correct the layer indexing and the typo.
  2. [Hypergraph Scanning & Flattening / Implementation] The notation for sequence length is inconsistent: Eq. (8) defines S^{(m)} of length N, while the implementation says H-ARW has fixed length T=0.7N and padding is applied for the remaining N-T nodes. Clarify whether N denotes the number of tiles or the maximum sequence length and where padding is inserted.
  3. [Experimental Settings] No information is given about tile extraction magnification, tissue masking, stain normalization, or the scanning and staining protocol of the private Prost dataset, which limits reproducibility and comparison. Please provide these details.
  4. [Tables 1 and 2] The tables do not report standard deviations or the number of runs; given the small accuracy differences among top methods, error bars are needed to support the claimed improvements over baselines.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the architecture is empirically evaluated rather than derived, self-citations are not load-bearing, and the efficiency concern about hyperedge construction is a scoping/factual issue, not a circularity.

full rationale

The paper does not claim a first-principles derivation; it proposes an architecture and evaluates it empirically. The HGMamba block is composed of hypergraph message passing (Eq. 7, citing Feng et al. 2019), bidirectional state space modeling (citing Gu and Dao 2023; Dao and Gu 2024), and the authors' own hypergraph scanning and flattening mechanism (Eqs. 8-9). These are design choices, not quantities fitted to the reported outcomes. The similarity-based adjacency construction (Eqs. 4-5) defines an input representation; it does not encode the final classification accuracy, so there is no self-definitional loop. The reported FLOPs and memory figures describe the HGMamba network, and the omission of the O(N^2 d) top-K similarity computation is a legitimate efficiency-scoping criticism, but it is a correctness or completeness issue, not a circular step. The ablation studies compare internal variants on the same benchmark datasets, which raises selection-bias concerns but does not make any 'prediction' reduce to a fitted parameter by construction. Self-citations (Lin et al. 2024, 2025; Liu et al. 2025; Lu et al. 2024; Yan et al. 2025; Gao et al. 2024b) appear in related-work listings and are not used to justify a central premise, a uniqueness theorem, or an ansatz. Accordingly, there is no load-bearing self-citation chain and no equation-level circularity; the appropriate score is 0.

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

The method relies on several hand-set hyperparameters and unchecked domain assumptions. No new natural entities are introduced. The most consequential hidden cost is the O(N^2) similarity search that underlies the claimed linear complexity, and the traversal-based sequence construction is assumed to preserve relational structure without proof.

free parameters (5)
  • Top-K neighbors K = 3
    Selected because the Prost ablation peaks at K=3 (Fig. 3c); used for all datasets without a theoretical justification.
  • Number of traversal sequences M = 8
    Hand-chosen in Implementations; no ablation or sensitivity analysis is reported for M.
  • Sequence length T = 0.7N
    Hand-chosen walk length for H-ARW in Implementations; no sensitivity analysis is reported.
  • Number of HGMamba layers = 2
    Selected based on Prost ablation (Fig. 3f) showing best F1 at 2 layers.
  • Tile size and tile encoder = 512x512, ResNet50
    Standard choices in WSI MIL, but they materially affect the features and downstream performance; no comparison to other tile sizes or encoders.
assumptions (5)
  • standard math Matrix operations, hypergraph convolution, Mamba/SSM equations, and Adam optimization are taken as valid background.
    The method relies on established equations (Eq. 7, SSM, normalization) without re-deriving them.
  • domain assumption Rule-based 4-neighbor adjacency and top-K cosine similarity capture meaningful tissue-level relationships.
    The hypergraph construction in the 'WSI Hypergraph Construction' section assumes these two similarity notions are sufficient and complementary for pathology.
  • domain assumption Flattening the hypergraph into traversal sequences preserves enough relational information for an SSM to model.
    The H-DFS and H-ARW strategies are proposed without a guarantee that sequences retain the high-order structure needed for accurate classification.
  • domain assumption Aggregating SSM outputs by averaging across sequences (Eq. 9) yields a faithful node representation for every tile.
    The aggregator assumes every node appears in at least one traversal sequence; with M=8 stochastic walks this is not guaranteed, and the paper does not analyze the effect of missing nodes.
  • domain assumption ABMIL pooling over tile embeddings is a sufficient slide-level classifier for these tasks.
    The paper uses ABMIL (Ilse et al. 2018) without comparing to other pooling strategies for the proposed architecture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hypergraph Mamba for Efficient Whole Slide Image Understanding." pith.science (2026). https://pith.science/paper/WG4IRA47

@misc{pith2026250517457,
  author       = {Pith},
  title        = {Pith review of: Hypergraph Mamba for Efficient Whole Slide Image Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WG4IRA47}},
  note         = {Machine review of arXiv:2505.17457}
}
read the original abstract

Whole Slide Images (WSIs) in histopathology pose a significant challenge for extensive medical image analysis due to their ultra-high resolution, massive scale, and intricate spatial relationships. Although existing Multiple Instance Learning (MIL) approaches like Graph Neural Networks (GNNs) and Transformers demonstrate strong instance-level modeling capabilities, they encounter constraints regarding scalability and computational expenses. To overcome these limitations, we introduce the WSI-HGMamba, a novel framework that unifies the high-order relational modeling capabilities of the Hypergraph Neural Networks (HGNNs) with the linear-time sequential modeling efficiency of the State Space Models. At the core of our design is the HGMamba block, which integrates message passing, hypergraph scanning & flattening, and bidirectional state space modeling (Bi-SSM), enabling the model to retain both relational and contextual cues while remaining computationally efficient. Compared to Transformer and Graph Transformer counterparts, WSI-HGMamba achieves superior performance with up to 7* reduction in FLOPs. Extensive experiments on multiple public and private WSI benchmarks demonstrate that our method provides a scalable, accurate, and efficient solution for slide-level understanding, making it a promising backbone for next-generation pathology AI systems.

Figures

Figures reproduced from arXiv: 2505.17457 by the authors.

Figure 1
Figure 1. Compared to existing relational and sequential [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall pipeline for WSI-HGMamba framework. WSIs are partitioned into tiles, and a pre-trained tile encoder [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Ablation experiments on the proposed WSI-HGMamba. (a) and (b) compare the computational costs (FLOPs) and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 38 canonical work pages

  1. [1]

    Behrouz, A.; and Hashemi, F. 2024. Graph mamba: Towards learning on graphs with state space models. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 119--130

  2. [2]

    H.; Cendra, F

    Chan, T. H.; Cendra, F. J.; Ma, L.; Yin, G.; and Yu, L. 2023. Histopathology whole slide image analysis with heterogeneous graph representation learning. In CVPR, 15661--15670

  3. [3]

    Chen, D.; Lin, Y.; Li, W.; Li, P.; Zhou, J.; and Sun, X. 2020. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In AAAI, volume 34, 3438--3445

  4. [4]

    J.; Lu, M

    Chen, R. J.; Lu, M. Y.; Shaban, M.; Chen, C.; Chen, T. Y.; Williamson, D. F.; and Mahmood, F. 2021 a . Whole slide images are 2d point clouds: Context-aware survival prediction using patch-based graph convolutional networks. In MICCAI, 339--349

  5. [5]

    J.; Lu, M

    Chen, R. J.; Lu, M. Y.; Weng, W.-H.; Chen, T. Y.; Williamson, D. F.; Manz, T.; Shady, M.; and Mahmood, F. 2021 b . Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In ICCV, 4015--4025

  6. [6]

    Dao, T.; and Gu, A. 2024. Transformers are SSM s: Generalized Models and Efficient Algorithms Through Structured State Space Duality. In ICML

  7. [7]

    Di, D.; Zhang, J.; Lei, F.; Tian, Q.; and Gao, Y. 2022 a . Big-hypergraph factorization neural network for survival prediction from whole slide image. IEEE Transactions on Image Processing, 31: 1149--1160

  8. [8]

    Di, D.; Zou, C.; Feng, Y.; Zhou, H.; Ji, R.; Dai, Q.; and Gao, Y. 2022 b . Generating hypergraph-based high-order representations of whole-slide histopathological images for survival prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5800--5815

Show all 47 references
  1. [9]

    Feng, Y.; You, H.; Zhang, Z.; Ji, R.; and Gao, Y. 2019. Hypergraph neural networks. In AAAI, volume 33, 3558--3565

  2. [10]

    Gao, C.; Sun, Q.; Zhu, W.; Zhang, L.; Zhang, J.; Liu, B.; and Zhang, J. 2024 a . Transformer based multiple instance learning for WSI breast cancer classification. Biomedical Signal Processing and Control, 89: 105755

  3. [11]

    Gao, Y.; Lu, J.; Li, S.; Li, Y.; and Du, S. 2024 b . Hypergraph-Based Multi-View Action Recognition Using Event Cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence

  4. [12]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv preprint arXiv:2312.00752

  5. [13]

    Han, X.; Zhou, H.; Tian, Z.; Du, S.; and Gao, Y. 2025. Inter-intra hypergraph computation for survival prediction on whole slide images. IEEE Transactions on Pattern Analysis and Machine Intelligence

  6. [14]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In CVPR, 770--778

  7. [15]

    Ilse, M.; Tomczak, J.; and Welling, M. 2018. Attention-based deep multiple instance learning. In ICML, 2127--2136

  8. [16]

    Keriven, N. 2022. Not too little, not too much: a theoretical analysis of graph (over) smoothing. NeurIPS, 35: 2268--2281

  9. [17]

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

  10. [18]

    Li, B.; Li, Y.; and Eliceiri, K. W. 2021. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In CVPR, 14318--14328

  11. [19]

    Li, H.; Zhang, Y.; Chen, P.; Shui, Z.; Zhu, C.; and Yang, L. 2024. Rethinking transformer for long contextual histopathology whole slide image analysis. arXiv preprint arXiv:2410.14195

  12. [20]

    Li, Z.; Cong, Y.; Chen, X.; Qi, J.; Sun, J.; Yan, T.; Yang, H.; Liu, J.; Lu, E.; Wang, L.; et al. 2023. Vision transformer-based weakly supervised histopathological image analysis of primary brain tumors. IScience, 26(1)

  13. [21]

    Lin, Y.; Lu, J.; Yong, Y.; and Zhang, J. 2025. MV-GMN: State Space Model for Multi-View Action Recognition. arXiv preprint arXiv:2501.13829

  14. [22]

    Lin, Y.; Zhang, J.; Li, S.; Xiao, J.; Xu, D.; Wu, W.; and Lu, J. 2024. Event USKT: U-State Space Model in Knowledge Transfer for Event Cameras. arXiv preprint arXiv:2411.15276

  15. [23]

    Liu, H.; Gong, Y.; Yan, Z.; Zhuang, Z.; and Lu, J. 2025. MSGM: A Multi-Scale Spatiotemporal Graph Mamba for EEG Emotion Recognition. arXiv preprint arXiv:2507.15914

  16. [24]

    Lu, J.; Yan, F.; Zhang, X.; Gao, Y.; and Zhang, S. 2024. Pathotune: Adapting visual foundation model to pathological specialists. In MICCAI, 395--406

  17. [25]

    Free Lunch

    Ren, R.; Li, Z.; and Liu, Y. 2024. Can Mamba Always Enjoy the" Free Lunch"? arXiv preprint arXiv:2410.03810

  18. [26]

    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. NeurIPS, 34: 2136--2147

  19. [27]

    Shi, J.; Shu, T.; Wu, K.; Jiang, Z.; Zheng, L.; Wang, W.; Wu, H.; and Zheng, Y. 2024 a . Masked hypergraph learning for weakly supervised histopathology whole slide image classification. Computer Methods and Programs in Biomedicine, 253: 108237

  20. [28]

    Shi, Z.; Zhang, J.; Kong, J.; and Wang, F. 2024 b . Integrative Graph-Transformer Framework for Histopathology Whole Slide Image Representation and Classification. In MICCAI, 341--350

  21. [29]

    J.; and Sinop, A

    Shirzad, H.; Velingker, A.; Venkatachalam, B.; Sutherland, D. J.; and Sinop, A. K. 2023. Exphormer: Sparse transformers for graphs. In ICML, 31613--31632

  22. [30]

    Tang, W.; Huang, S.; Zhang, X.; Zhou, F.; Zhang, Y.; and Liu, B. 2023. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In ICCV, 4078--4087

  23. [31]

    Wang, X.; Yang, S.; Zhang, J.; Wang, M.; Zhang, J.; Yang, W.; Huang, J.; and Han, X. 2022. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis, 81: 102559

  24. [32]

    Weitz, P.; Wang, Y.; Hartman, J.; and Rantalainen, M. 2021. An investigation of attention mechanisms in histopathology whole-slide-image analysis for regression objectives. In ICCV, 611--619

  25. [33]

    Xing, Z.; Ye, T.; Yang, Y.; Liu, G.; and Zhu, L. 2024. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In MICCAI, 578--588

  26. [34]

    Xu, H.; Usuyama, N.; Bagga, J.; Zhang, S.; Rao, R.; Naumann, T.; Wong, C.; Gero, Z.; Gonz \'a lez, J.; Gu, Y.; et al. 2024. A whole-slide foundation model for digital pathology from real-world data. Nature, 1--8

  27. [35]

    Yan, F.; Wu, J.; Li, J.; Wang, W.; Lu, J.; Chen, W.; Gao, Z.; Li, J.; Yan, H.; Ma, J.; et al. 2025. PathOrchestra: A Comprehensive Foundation Model for Computational Pathology with Over 100 Diverse Clinical-Grade Tasks. arXiv preprint arXiv:2503.24345

  28. [36]

    Yang, S.; Wang, Y.; and Chen, H. 2024. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. In MICCAI, 296--306

  29. [37]

    Yao, J.; Zhu, X.; Jonnagaddala, J.; Hawkins, N.; and Huang, J. 2020. Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis, 65: 101789

  30. [38]

    Yu, M.; Wang, H.; Fu, X.; Gao, J.; Liu, Z.; and Li, X. 2024. DualGCN-MIL: Whole slide image classification based on double relationship graph learning. In ICASSP, 1986--1990

  31. [39]

    Yu, W.; and Wang, X. 2024. MambaOut: Do We Really Need Mamba for Vision? arXiv preprint arXiv:2405.07992

  32. [40]

    T.; Han, X.; Trinh, V

    Zhang, J.; Nguyen, A. T.; Han, X.; Trinh, V. Q.-H.; Qin, H.; Samaras, D.; and Hosseini, M. S. 2025. 2dmamba: Efficient state space model for image representation with applications on giga-pixel whole slide image classification. In CVPR, 3583--3592

  33. [41]

    Zhang, Y.; Li, H.; Sun, Y.; Zheng, S.; Zhu, C.; and Yang, L. 2024. Attention-challenging multiple instance learning for whole slide image classification. In ECCV, 125--143

  34. [42]

    Zhao, Y.; Yang, F.; Fang, Y.; Liu, H.; Zhou, N.; Zhang, J.; Sun, J.; Yang, S.; Menze, B.; Fan, X.; et al. 2020. Predicting lymph node metastasis using histopathological images based on multiple instance learning with deep graph convolution. In CVPR, 4837--4846

  35. [43]

    H.; Green, E

    Zheng, Y.; Gindra, R. H.; Green, E. J.; Burks, E. J.; Betke, M.; Beane, J. E.; and Kolachalama, V. B. 2022. A graph-transformer for whole slide image classification. IEEE Transactions on Medical Imaging, 41(11): 3003--3015

  36. [44]

    Zheng, Y.; Jiang, Z.; Xie, F.; Shi, J.; Zhang, H.; Huai, J.; Cao, M.; and Yang, X. 2020. Diagnostic regions attention network (dra-net) for histopathology wsi recommendation and retrieval. IEEE Transactions on Medical Imaging, 40(3): 1090--1103

  37. [45]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  39. [47]

    !1A Qa

    aaai2026 document ablation_experiments_crop.pdf0000664000000000000000000113324315044067041015520 0ustar rootroot 4 0 obj << /Length 46 /Filter /FlateDecode >> stream x+2T0BC ]C =CsK RP!9K3P endstream endobj 1 0 obj << /Type /XObject /Subtype /Form /FormType 1 /PTEX.FileName (....

Pith tools

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