Pith. sign in

REVIEW 2 major objections 5 minor 57 references

SAC-MIL is an all-MLP model that encodes patch positions through polar-coordinate rotation and correlates every patch in a whole slide in linear time, reporting the best accuracy, AUC, and F1 on three histopathology benchmarks.

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 ↗

A spatial-aware MLP-Mixer with polar rotary position embedding and channel-shifting SAC blocks achieves state-of-the-art slide-level classification on CAMELYON-16, TCGA-LUNG, and TCGA-BRCA.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A clever all-MLP WSI classifier with a real design idea, but the central full-correlation claim is undermined by a math-versus-pseudo-code contradiction that needs to be resolved before the numbers can be trusted. the 2 major comments →

arxiv 2509.03973 v1 pith:GEPTG72P submitted 2025-09-04 cs.CV cs.AI

SAC-MIL: Spatial-Aware Correlated Multiple Instance Learning for Histopathology Whole Slide Image Classification

Yu Bai , Zitong Yu , Haowen Tian , Xijing Wang , Shuo Yan , Lin Wang , Honglin Li , Xitong Ling
show 6 more authors
Bo Zhang Zheng Zhang Wufan Wang Hui Gao Xiangyang Gong Wendong Wang
This is my paper
classification cs.CV cs.AI
keywords whole slide image classificationmultiple instance learningMLP-Mixerrotary position embeddingpolar coordinateslinear complexityhistopathologyweakly supervised learning
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 reading

Whole-slide cancer classification usually treats each slide as a bag of tens of thousands of patches: linear-time attention-free models ignore relations between patches, while Transformer models capture them at quadratic cost or need specialized GPU kernels. SAC-MIL is an attempt to get both benefits from MLP layers alone, through a positional module (PROPE) that rotates patch features using normalized polar coordinates instead of sequence indices, and a SAC block that cyclically shifts channel slices between instances and mixes them with channel-wise MLPs to perform what the paper calls full instance correlation. The paper argues that three SAC blocks, with region sizes growing from 64 to 4096 to 262,144, push the effective context to the whole slide at linear cost in sequence length, and reports state-of-the-art accuracy, AUC, and F1 on CAMELYON-16, TCGA-LUNG, and TCGA-BRAC with both ResNet50 and UNI features. If the mechanism works as specified, high-performance WSI classification no longer depends on custom attention kernels, easing deployment in clinics.

Core claim

On its own terms, the paper claims MLP-based multiple instance learning can match or beat Transformer-based methods on whole-slide classification. First, PROPE normalizes each patch's 2D coordinates per slide, converts them to polar radius and angle, and rotates features by e^{i(ρθ+α)}; per-slide normalization maps training and test sequences of different lengths into the same range, which the paper says removes length-extrapolation failure. Second, the SAC block splits instance features into channel folds, rolls each fold across instances by an increasing step (partial correlation), mixes with a channel-wise MLP, reverses the rolls, and mixes again (full correlation); region size grows expo

What carries the argument

The load-bearing object is the SAC block, an MLP-only correlation primitive: each instance is split along the channel dimension into folds; each fold is cyclically shifted across instances by an amount that grows with the fold index (the partial-correlation step); a channel-wise MLP mixes the shifted folds; the rolls are undone and a second channel-wise MLP mixes again (the full-correlation step). Working alongside it is PROPE, a parameter-free positional embedding that rotates each patch feature by a phase built from its normalized polar coordinates (radius ρ, angle α) rather than its slot in the input sequence. FPS+KNN region partitioning feeds the blocks local neighborhoods, and the regio

Load-bearing premise

Full correlation within a region requires as many channel folds per instance as there are instances in the region; the main text (8 folds) and the appendix pseudo code (64 folds) disagree on this count, and the paper never resolves which one is implemented.

What would settle it

Run the Fig. 5 effective-context test on the released SAC-MIL code with an input of length 65,536 and region sizes 64/4096/262,144, or simply count the channel folds in the shift function. If instances are split into D/k = 8 folds, one block correlates only 8 neighbours and the number of outputs that change when the middle input is zeroed will be far below the sequence length; if the appendix's k = 64 folds are implemented, all outputs change. Either observation settles the specification.

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

If this is right

  • If the full-correlation mechanism holds, whole-slide classifiers reach Transformer-comparable accuracy with pure MLP operations, so deployment no longer depends on custom CUDA kernels or specific GPU architectures.
  • Per-slide coordinate normalization makes the positional embedding robust to slides of very different sizes, so a model trained on short sequences can be applied to longer ones without retraining.
  • Linear complexity with respect to instance count means a stack of three blocks can process typical slides with over 10,000 patches without subsampling.
  • Because the SAC block is length-agnostic, the same architecture transfers to other bag-of-instance pathology tasks the paper names as future work, such as survival prediction.

Where Pith is reading between the lines

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

  • The full-correlation claim is only as strong as the fold count: the main-text equations imply D/k = 8 channel folds per instance (for D=512, k=64), which after three blocks would limit the effective context to roughly 512 instances, whereas the appendix pseudo code splits into k = 64 folds, which supports the claimed context of 64³ = 262,144; the paper never flags the difference.
  • The normalized-polar rotation is geometry-agnostic, so the same positional encoding could be dropped into other tile- or point-based medical imaging tasks, such as spatial transcriptomics or gigapixel scanning, wherever coordinates are known.
  • The Fig. 5 zeroed-instance test is a cheap, decisive experiment: run it on released code and count the changed outputs to determine which fold count is actually implemented.
  • Since SAC blocks are drop-in MLP replacements for self-attention, they invite head-to-head benchmarking against Transformer blocks in multimodal pathology settings with no inference-stack changes.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 5 minor

Summary. This paper proposes SAC-MIL, an all-MLP multiple instance learning model for whole-slide histopathology classification. It partitions patch instances into spatial regions via FPS/KNN, applies a polar rotary position embedding (PROPE) based on normalized coordinates, and stacks SAC blocks that shift channel folds across instances and apply channel-wise MLPs to correlate instances in linear time. The authors report state-of-the-art accuracy/AUC/F1 on CAMELYON-16, TCGA-LUNG, and TCGA-BRCA with ResNet-50 and UNI features, and present ablations of PROPE, correlation methods, and an effective-context-length (ECL) test against Cycle-MLP and Chord-Mixer.

Significance. If the core mechanism works as stated, SAC-MIL is an attractive practical alternative to transformer-based MIL: linear complexity in sequence length, no custom CUDA kernels, and strong performance on standard benchmarks. The paper's all-MLP design and the empirical ECL comparison are useful and clearly presented. However, the central claim of full instance correlation is not currently pinned down, because the formal shift equation in Sec. III-D and the pseudo-code in Appendix I specify different schedules. The contribution cannot be fully assessed until this ambiguity is resolved.

major comments (2)
  1. [Sec. III-D, Eq. (10), Appendix I] Eq. (10) and Algorithm 2 specify different shift schedules. In Eq. (10), t_j is an absolute channel index; with D=512 and k=64, floor(t_j/k) takes only the eight values 0,...,7. Thus the first SAC block (region size 64, l=0) shifts folds by at most 7, so a 64-instance region cannot be fully correlated; the same shortfall propagates to larger regions. Algorithm 2, by contrast, splits the channel dimension into k=64 chunks and shifts chunk idx by idx*scale (idx=0,...,63), which is the schedule that supports the claimed exponential growth. If Algorithm 2 is the implementation, Eq. (10) should use floor(t_j/(D/k)) (equivalently floor(t_j*k/D)); if Eq. (10) is the implementation, the full-correlation claim and the ECL results in Figs. 4-6 are unsupported. The manuscript never flags this contradiction and no code is provided. Because every downstream claim (linear-time full correlation, SOTA r
  2. [Sec. IV-B, Figs. 4-6] The ECL test is presented as evidence that SAC-MIL performs full instance correlations, but the protocol zeroes one middle instance and counts changed outputs. This is only a necessary condition for full correlation and is not a proof that all instance pairs interact. More importantly, the interpretation of the test changes completely under the two shift schedules in Eq. (10) vs Algorithm 2. Please provide a formal statement of the correlation property for the exact implemented schedule, or release the code so that the schedule can be verified. The current empirical ECL curves do not resolve the Eq. (10)/Algorithm 2 ambiguity.
minor comments (5)
  1. [Sec. III-D, Eq. (9)] The right-hand side of Eq. (9) uses \hat{h}_{i,t_j}; as written it defines a value in terms of itself. It should be h_{i,t_j} or otherwise define the shifted tensor explicitly.
  2. [Sec. IV, Model] The third region size is given as 262114. Since 64^3 = 262144, this is likely a typo and should be corrected.
  3. [Appendix III, Table VI] The PROPE row in Table VI is identical to Table V (90.4/95.0/90.1) yet is labeled TCGA-BRAC. This is impossible given the other rows in the table and is probably a copy-paste error.
  4. [Sec. III-C and Conclusion] The text refers to the 'length exploitation issue' and the 'length exploration issue'; these should be 'length extrapolation issue'.
  5. [Appendix I, Algorithm 2] The pseudo-code is labeled Algorithm 2 but the text says 'pseudo code ... illustrated in 1'. Also, the check 'if x.shape[-2] == 0' appears to be intended as a length-zero check and should be 'if x.shape[-2] == 0' or similar with the correct dimension. Please clarify.

Circularity Check

0 steps flagged

No derivation reduces to its inputs; only a minor non-load-bearing self-citation and a correctness ambiguity, not circularity.

full rationale

The central claims of SAC-MIL are constructional and empirical. PROPE is a coordinate-based positional encoding with a chosen scaling factor; the SAC block's full-correlation property is supported by direct ECL perturbation experiments (Figs. 4-6) on random inputs, which is a behavioral check rather than an assumed conclusion. The reported SOTA results are benchmark comparisons against published baselines, not quantities derived from the method's own definitions. The only self-citation, Norma [23], is used as a comparison baseline and is not load-bearing for the SAC-MIL derivation. Appendix IV's selection of lambda on the same evaluation datasets is a hyperparameter-selection concern, not an identity or reduction of the derivation. A genuine correctness issue exists: Eq. 10 specifies D/k folds (8 for D=512, k=64), while Algorithm 2 splits into k chunks (64), so the claimed full-correlation property is ambiguous; however, this is an internal inconsistency, not circularity. Overall circularity is minimal.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central architecture relies on a small set of hand-set hyperparameters, lambda, region sizes, N, and the fold count. The fold count is ambiguous between equations and pseudo code. No new physical entities are introduced. The main hidden assumptions are the comparability of baselines and the validity of the ECL proxy.

free parameters (4)
  • lambda (PROPE radius scaling) = 512
    Selected empirically on the same test folds in Appendix IV, Table VII; controls the polar radius range.
  • Region sizes k_l = 64, 4096, 262144
    Set by hand to grow exponentially; no systematic search or sensitivity analysis is reported.
  • Number of SAC blocks N = 3
    Set in the model configuration; the ECL grows with N, but no ablation over N is given.
  • Number of channel folds = 64 (pseudo code) or 8 (Eq. 10)
    The full-correlation property depends on this value, and the paper gives contradictory definitions in Section III-D and Appendix I.
axioms (4)
  • domain assumption Baseline results are directly comparable across methods despite potential differences in data splits, preprocessing, and hyperparameters.
    The paper reports five-fold cross-validation for its own runs but does not state whether baseline numbers were re-run under the same folds or taken from original papers; the SOTA claim depends on this comparability.
  • ad hoc to paper The ECL test, setting one instance to zero and counting changed outputs, is a valid proxy for full instance correlation.
    Used in Section IV-B and Figures 4 to 6 to validate the central full-correlation claim; this is an operational definition not derived from the architecture.
  • domain assumption FPS and KNN produce regions that preserve local tissue context, and arranging region instances contiguously in the sequence is lossless.
    Section III-B introduces this ordering; the shift-based correlation depends on it, and no sensitivity analysis is given.
  • domain assumption Patch coordinates are available as top-left corners for all tissue patches and are used without further preprocessing details.
    Required by PROPE and region clustering; the paper does not specify tissue segmentation or coordinate extraction steps.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SAC-MIL: Spatial-Aware Correlated Multiple Instance Learning for Histopathology Whole Slide Image Classification." pith.science (2026). https://pith.science/paper/GEPTG72P

@misc{pith2026250903973,
  author       = {Pith},
  title        = {Pith review of: SAC-MIL: Spatial-Aware Correlated Multiple Instance Learning for Histopathology Whole Slide Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GEPTG72P}},
  note         = {Machine review of arXiv:2509.03973}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose Spatial-Aware Correlated Multiple Instance Learning (SAC-MIL) for performing WSI classification. SAC-MIL consists of a positional encoding module to encode position information and a SAC block to perform full instance correlations. The positional encoding module utilizes the instance coordinates within the slide to encode the spatial relationships instead of the instance index in the input WSI sequence. The positional encoding module can also handle the length extrapolation issue where the training and testing sequences have different lengths. The SAC block is an MLP-based method that performs full instance correlation in linear time complexity with respect to the sequence length. Due to the simple structure of MLP, it is easy to deploy since it does not require custom CUDA kernels, compared to Transformer-based methods for WSI classification. SAC-MIL has achieved state-of-the-art performance on the CAMELYON-16, TCGA-LUNG, and TCGA-BRAC datasets. The code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2509.03973 by Bo Zhang, Haowen Tian, Honglin Li, Hui Gao, Lin Wang, Shuo Yan, Wendong Wang, Wufan Wang, Xiangyang Gong, Xijing Wang, Xitong Ling, Yu Bai, Zheng Zhang, Zitong Yu.

Figure 1
Figure 1. Figure 1: (a) MLP-based methods employ MLPs with vanilla attention (e.g., Sigmoid function) to perform WSI classification. These methods exhibit linear complexity with respect to the instance number but do not explicitly model correlations between instances. (b) Transformer￾based methods utilize self-attention mechanisms with quadratic com￾plexity to capture correlations between instances and also encode positional … view at source ↗
Figure 2
Figure 2. Figure 2: We first divide all instances into multiple regions using [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall Architecture of SAC-MIL. SAC-MIL uses FPS and KNN to split instances into multiple regions, indicated by different colors. PROPE encodes positional information based on spatial coordinates. Each SAC block first splits each instance into multiple folds along the channel dimension, shifts each fold with increased steps, and finally applies a channel-wise MLP to perform partial instance correlations. … view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of two stacked SAC blocks. (a) The first SAC block. (b) The second SAC block. Partial correlation makes each instance encode partial information of other instances, which can be seen as fold-level correlation. Full correlation enables each instance to encode full information of other instances; each fold contains all the information from other folds within the same region. As the number of SAC… view at source ↗
Figure 5
Figure 5. Figure 5: Number of instances changes for three methods after setting the middle instance to a zero vector. (a), (b), (c), and (d) represent the WSI sequence lengths of 512, 4096, 16384, and 65536, respectively. compared to Flash-Attn [51], and Transformer-based meth￾ods do not significantly outperform MLP-Mixer based meth￾ods, indicating that under small training sample settings, Transformer-based methods may not h… view at source ↗
Figure 4
Figure 4. Figure 4: The L2 norm of the differences between the outputs from two forward passes is calculated. The input sequence length is set to 10,000. The x-axis represents the instance index, and the y-axis represents the L2 norm value. (a) Cycle-MLP, (b) Chord-Mixer, and (c) SAC-MIL. Methods Accuracy AUC F1 Cycle-MLP [9] 89.51.5 92.53.5 89.01.8 Chord-Mixer [10] 90.01.4 93.00.5 89.51.2 Nystrom [37] ¨ 88.51.6 92.00.4 88.01… view at source ↗
Figure 6
Figure 6. Figure 6: The number of affected instances as the number of layers increases for three methods. The length of the input sequence is set to 10,000. The x-axis represents the number of layers, and the y￾axis represents the number of affected instances. The three methods evaluated are Cycle-MLP, Chord-Mixer, and SAC-MIL [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Patch visualization produced by SAC-MIL. Images in the left column is the ground truth of tumor region (red). Left Column:Ground truth images of the tumor regions, highlighted in red. Right Column: The L2 norm of each output instance. block can perform full instance correlation and encode posi￾tional information with only linear computational complexity, and the simple MLP-style structure making it friendl… view at source ↗
Figure 8
Figure 8. Figure 8: Patch visualization produced by SAC-MIL. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. [17] Richard J. Chen et al. Scaling vision transformers to gigapixel im￾ages via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. [18] Alexey Dosovitskiy et al. An image is worth 1… 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

57 extracted references · 47 canonical work pages · 4 internal anchors

  1. [1]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    work page 2016

  2. [2]

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

    Zhuchen Shao et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. In Advances in Neural Information Processing Systems (NeurIPS) , volume 34, pages 2136–2147, 2021

    work page 2021

  3. [3]

    Feature re-embedding: Towards foundation model- level performance in computational pathology

    Wenhao Tang et al. Feature re-embedding: Towards foundation model- level performance in computational pathology. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  4. [4]

    Conditional positional encodings for vision transformers

    Xiangxiang Chu et al. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882 , 2021

    Pith/arXiv arXiv 2021

  5. [5]

    Roformer: Enhanced transformer with rotary position embedding

    Jianlin Su et al. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024

    work page 2024

  6. [7]

    Unified-IO 2: Scaling autoregressive multimodal models with vision language audio and action

    Jiasen Lu et al. Unified-IO 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  7. [9]

    CycleMLP: A MLP-like Architecture for Dense Prediction

    Shoufa Chen et al. Cyclemlp: A mlp-like architecture for dense prediction. arXiv preprint arXiv:2107.10224 , 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  8. [10]

    ChordMixer: A Scalable Neural Attention Model for Sequences with Different Lengths

    Ruslan Khalitov et al. Chordmixer: A scalable neural attention model for sequences with different lengths. arXiv preprint arXiv:2206.05852 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  9. [11]

    Metaformer is actually what you need for vision

    Weihao Yu et al. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  10. [12]

    Shapley values-enabled progressive pseudo bag augmentation for whole-slide image classification

    Renao Yan et al. Shapley values-enabled progressive pseudo bag augmentation for whole-slide image classification. IEEE Transactions on Medical Imaging , 2024

    work page 2024

  11. [13]

    Patch-based convolutional neural network for whole slide tissue image classification

    Le Hou et al. Patch-based convolutional neural network for whole slide tissue image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016

    work page 2016

  12. [14]

    Eliceiri

    Bin Li, Yin Li, and Kevin W. Eliceiri. Dual-stream multiple in- stance 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 (CVPR), 2021

    work page 2021

  13. [15]

    Lu et al

    Ming Y . Lu et al. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering , 5(6):555–570, 2021

    work page 2021

  14. [16]

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

    Hongrun Zhang et al. Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. 10 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. XX, NO. XX, XXXX 2020 Methods CAMELYON-16 TCGA-LUNG TCGA-BRAC ACC AUC F1 ACC AUC F1 ACC AUC F1 λ = 1 89 .9 1 .3 91 .0 1 .3 89 .7 0 .3 94 .8 1 .1 97 .4 0 .5 95 .2 1 .0 ...

    work page 2020

  15. [17]

    Chen et al

    Richard J. Chen et al. Scaling vision transformers to gigapixel im- ages via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  16. [18]

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

    Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 , 2020

    Pith/arXiv arXiv 2010

  17. [19]

    Self-attention with relative position representations

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018

    Pith/arXiv arXiv 2018

  18. [20]

    Transformer-xl: Attentive language models beyond a fixed-length context

    Zihang Dai et al. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860 , 2019

    Pith/arXiv arXiv 1901

  19. [21]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc ¸ois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, pages 5156–5165, 2020

    work page 2020

  20. [22]

    A whole-slide foundation model for digital pathology from real-world data

    Hanwen Xu et al. A whole-slide foundation model for digital pathology from real-world data. Nature, pages 1–8, 2024

    work page 2024

  21. [23]

    Norma: A noise robust memory-augmented framework for whole slide image classification

    Yu Bai et al. Norma: A noise robust memory-augmented framework for whole slide image classification. In European Conference on Computer Vision (ECCV). Springer, Cham, 2025

    work page 2025

  22. [24]

    Vaswani et al

    A. Vaswani et al. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS) , 2017

    work page 2017

  23. [25]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher R ´e. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021

    Pith/arXiv arXiv 2021

  24. [26]

    Sparse factorization of square matrices with application to neural attention modeling

    Ruslan Khalitov et al. Sparse factorization of square matrices with application to neural attention modeling. Neural Networks , 152:160– 168, 2022

    work page 2022

  25. [27]

    Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathol- ogy

    Shu Yang, Yihui Wang, and Hao Chen. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathol- ogy. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). Springer Nature Switzerland, 2024

    work page 2024

  26. [28]

    MamMIL: Multiple Instance Learning for Whole Slide Images with State Space Models

    Zijie Fang et al. Mammil: Multiple instance learning for whole slide images with state space models. arXiv preprint arXiv:2403.05160, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [29]

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

    Gabriele Campanella et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine, 25(8):1301–1309, 2019

    work page 2019

  28. [30]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 , 2023

    Pith/arXiv arXiv 2023

  29. [31]

    Structured state space models for multiple instance learning in digital pathology

    Leo Fillioux et al. Structured state space models for multiple instance learning in digital pathology. In International Conference on Medi- cal Image Computing and Computer-Assisted Intervention (MICCAI) . Springer Nature Switzerland, 2023

    work page 2023

  30. [32]

    The farthest point strategy for progressive image sampling

    Yuval Eldar et al. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing , 6(9):1305–1315, 1997

    work page 1997

  31. [33]

    Rotary position embedding for vision transformer

    Byeongho Heo et al. Rotary position embedding for vision transformer. In European Conference on Computer Vision (ECCV) . Springer, Cham, 2025

    work page 2025

  32. [34]

    Fit: Flexible vision transformer for diffusion model

    Zeyu Lu et al. Fit: Flexible vision transformer for diffusion model. arXiv preprint arXiv:2402.12376 , 2024

    Pith/arXiv arXiv 2024

  33. [35]

    Tolstikhin et al

    Ilya O. Tolstikhin et al. Mlp-mixer: An all-mlp architecture for vision. In Advances in Neural Information Processing Systems (NeurIPS) , volume 34, pages 24261–24272, 2021

    work page 2021

  34. [36]

    Attention-based deep multiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International Conference on Machine Learning (ICML). PMLR, 2018

    work page 2018

  35. [37]

    Nystromformer: A nystrom-based algorithm for approximating self-attention

    Yunyang Xiong et al. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, number 16, 2021

    work page 2021

  36. [38]

    Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis

    Honglin Li et al. Rethinking transformer for long contextual histopathol- ogy whole slide image analysis. arXiv preprint arXiv:2410.14195, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [39]

    Longnet: Scaling transformers to 1,000,000,000 tokens

    Jiayu Ding et al. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486 , 2023

    Pith/arXiv arXiv 2023

  38. [40]

    A length-extrapolatable transformer

    Yutao Sun et al. A length-extrapolatable transformer. arXiv preprint arXiv:2212.10554, 2022

    Pith/arXiv arXiv 2022

  39. [42]

    Pay attention to mlps

    Hanxiao Liu et al. Pay attention to mlps. In Advances in Neural Information Processing Systems (NeurIPS) , volume 34, pages 9204– 9215, 2021

    work page 2021

  40. [43]

    Deformable convolutional networks

    Jifeng Dai et al. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2017

    work page 2017

  41. [44]

    Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classifi- cation

    Honglin Li et al. Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classifi- cation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2023

    work page 2023

  42. [45]

    Multiple instance learning framework with masked hard instance mining for whole slide image classification

    Wenhao Tang et al. Multiple instance learning framework with masked hard instance mining for whole slide image classification. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2023

    work page 2023

  43. [46]

    Histopathology whole slide image analysis with heterogeneous graph representation learning

    Tsai Hor Chan et al. Histopathology whole slide image analysis with heterogeneous graph representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. YU BAI et al.: SAC-MIL: SPATIAL-AWARE CORRELATED MULTIPLE INSTANCE LEARNING FOR HISTOPATHOLOGY WHOLE SLIDE IMAGE CLASSIFICATION 11

    work page 2023

  44. [47]

    A graph-transformer for whole slide image classification

    Yi Zheng et al. A graph-transformer for whole slide image classification. IEEE Transactions on Medical Imaging , 41(11):3003–3015, 2022

    work page 2022

  45. [48]

    Chen et al

    Richard J. Chen et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850–862, 2024

    work page 2024

  46. [49]

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

    Babak Ehteshami Bejnordi et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA, 318(22):2199–2210, 2017

    work page 2017

  47. [50]

    Smith, and Mike Lewis

    Ofir Press, Noah A. Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. In ICLR, 2021

    work page 2021

  48. [51]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao et al. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 16344–16359, 2022

    work page 2022

  49. [52]

    Resmlp: Feedforward networks for image classifica- tion with data-efficient training

    Hugo Touvron et al. Resmlp: Feedforward networks for image classifica- tion with data-efficient training. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(4):5314–5321, 2022

    work page 2022

  50. [53]

    Hanxiao Liu, Zihang Dai, David So, and Quoc V . Le. Pay attention to mlps. Advances in Neural Information Processing Systems , 34:9204– 9215, 2021

    work page 2021

  51. [54]

    Vision permutator: A permutable mlp-like architecture for visual recognition

    Qibin Hou et al. Vision permutator: A permutable mlp-like architecture for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1328–1334, 2022

    work page 2022

  52. [55]

    An image patch is a wave: Phase-aware vision mlp

    Yehui Tang et al. An image patch is a wave: Phase-aware vision mlp. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10935–10944, 2022

    work page 2022

  53. [56]

    Global filter networks for image classification

    Yongming Rao et al. Global filter networks for image classification. Advances in Neural Information Processing Systems, 34:980–993, 2021

    work page 2021

  54. [57]

    Extending context window of large language models via positional interpolation

    Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595 , 2023

    Pith/arXiv arXiv 2023

  55. [58]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision, pages 1026–1034, 2015

    work page 2015

  56. [59]

    COVID-19 pandemic and managing supply chain risks: NVIDIA’s graphics card shortage case analysis, 2021

    N ´andor Hajdu. COVID-19 pandemic and managing supply chain risks: NVIDIA’s graphics card shortage case analysis, 2021

    work page 2021

  57. [60]

    NVIDIA’s bet on artificial intelligence: The impact on healthcare

    Daniela da Silva Ferreira Fernandes. NVIDIA’s bet on artificial intelligence: The impact on healthcare. Master’s thesis, Universidade NOV A de Lisboa (Portugal), 2024

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.