acceptodds
Under review as a conference paper at ICLR 2027

One Tile, Multiple Instances: Rethinking MIL for Sparse Diagnostic Evidence

10 views

Abstract

In weakly supervised Whole Slide Image (WSI) classification, feature extractors typically compress each image tile into a single global embedding. Consequently, slide-level aggregators are restricted to this coarse tile scale, concealing fine-grained sub-tile evidence from the attention mechanism. We introduce DI-MIL, a framework that decouples encoding context from instance granularity through decomposed instances. By clustering dense spatial tokens from a frozen foundation model within each tile, DI-MIL converts a single tile into multiple independently weighted instance embeddings. As a training-free post-encoding module, DI-MIL integrates seamlessly into existing pipelines without requiring re-encoding or downstream architectural modifications. We evaluate DI-MIL on cytopathology, a challenging testbed where sparse diagnostic signals are easily diluted within standard tiles. Across four datasets, three frozen foundation models, and two attention-based aggregators, DI-MIL demonstrates consistent efficacy, improving 67 of 72 metric-level comparisons, with the largest mean gains reaching 3.64 points under cytopathology-specific backbones. In a broader comparison against seven representative MIL baselines, DI-MIL paired with ACMIL achieves highest mean performance in 33 of 36 backbone-dataset-metric comparisons. Ablations show that direct smaller tiling inflates the extracted tile count by up to 43.3 with non-monotonic performance, whereas DI-MIL incurs zero additional image-extraction overhead while achieving the strongest overall results. These results establish instance construction as an orthogonal design dimension in MIL, supporting DI-MIL as a cost-efficient solution under sparse diagnostic evidence. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multiple Instance Learning, Computational Pathology, Cytopathology, Whole-Slide Image Analysis, Weakly Supervised Learning, Instance Granularity, Instance Decomposition, Sparse Diagnostic Evidence

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NzUBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To tackle Whole Slide Image (WSI) classification under weak slide-level annotations, the deep learning community has widely adopted Multiple Instance Learning (MIL) frameworks. Standard MIL pipelines crop a WSI into small tiles, extract representations using a pathology-pretrained Feature Extractor (FE), and aggregate these embeddings for final slide-level prediction. However, relying exclusively on the FE's [CLS] token causes oversmoothing of intra-tile context, introducing significant limitations in cytopathology where diagnostic evidence appears sparsely.To address this issue, the paper propose DI-MIL, a novel approach that preserves and utilizes the FE's patch token features rather than discarding them. DI-MIL clusters patch features within each tile to retain fine-grained spatial representations. Since the ViT-based FE inherently partitions tiles into sub-patches, DI-MIL requires no additional tile pre-processing. The newly enriched bag can be seamlessly integrated as a plug-and-play module into existing MIL backbones (evaluated with AB-MIL and ACMIL in this work). Experimental results demonstrate that DI-MIL effectively overcomes the limitations of naive [CLS]-only baselines, exhibiting superior performance in cytopathology tasks.

Critical strengths
  1. The core idea of leveraging both the [CLS] token and patch tokens effectively tackles the common practice of discarding fine-grained intra-tile information. Furthermore, because the Vision Transformer (ViT) encoder inherently partitions tiles into sub-patches during its internal operations, the proposed method recovers detailed local representations without requiring any additional computational overhead or manual tile pre-processing.

  2. The manuscript addresses cytopathology and whole-slide image (WSI) analysis, a domain that has received relatively less attention compared to histopathology. Given the distinct diagnostic challenges of cytopathology (such as sparsely distributed cellular features), this work addresses a crucial gap and holds strong potential for broader impact within the digital pathology community.

  3. (Assuming my understanding is correct) The proposed DI-MIL framework operates as a plug-and-play module, demonstrating high adaptability and scalability. Since the vast majority of modern pathology feature extractors (FEs) rely on ViT-based architectures, the clustering approach can be seamlessly integrated across various pre-trained encoders without requiring retrained or specialized feature extraction pipelines.

Critical weaknesses
  1. Merging WSI tiles to capture fine-grained representations is an established methodology in the literature, rendering the proposed method somewhat incremental. To the best of my knowledge, existing approaches in this direction fall into two primary paradigms.

    1-A. Hierarchical approaches:

    • HIPT [1] progressively refines representations by aggregating features from fine-grained to coarse-grained tile levels, preventing uniform information loss. Contrary to the claim in the manuscript that only a single attention score is assigned to a patch, class tokens across resolutions receive complex, contextualized attention scores via self-attention.

    • Beyond this, numerous studies have leveraged multi-resolution strategies to mitigate the oversmoothing of information that occurs during single-patch feature extraction [2-7].

    1-B. clustering

    • Multiple studies have already incorporated clustering for instance selection and aggregation within MIL [8-10].

    • Despite this prior work, the clear distinction between existing clustering-based approaches and the proposed method remains insufficiently articulated. Furthermore, such clustering-based approaches should be included as direct baselines to evaluate competitive performance.

    • While the baselines presented in Table 2 are valid, they stem from fundamentally different motivations. For instance, MambaMIL emphasizes sequential modeling, whereas DTFD-MIL focuses on feature expansion via pseudo-bag construction. Consequently, relying solely on these models is insufficient to thoroughly evaluate the specific motivation driving the proposed method.

  2. Regarding the claimed limitation of existing aggregation methods:

  • It is acknowledged that early Multiple Instance Learning approaches relied on relatively naive aggregation schemes [11-12].

  • However, recent studies cannot be simplified as merely applying a single attention mechanism to collapse representations into a single vector. Hierarchical graph-based methods utilize complex message passing and multimodal fusion [13-15]. Furthermore, recent work on MIL in hyperbolic space expands representation capacity by projecting fine-grained information along geodesics [16].

  • A primary concern is that the argument presented in the manuscript appears to generalize the limitations of a narrow subset of MIL methods as if they apply to the entire paradigm.

  • Following the same line of logic, the fundamental mechanism of the proposed method remains unchanged: it still collapses a slightly more complex bag into a single vector using a single attention operation (of AB-MIL and ACMIL).

  1. According to Table 8, the training seed is set strictly to 1024, indicating that the experiments were conducted deterministically in a single run. This setup introduces two primary concerns:

    3-A. It is unclear how the reported standard deviations were calculated, as a single deterministic run cannot produce a standard deviation. Furthermore, clarification is needed on whether all baseline methods were trained under this exact single-seed configuration.

    3-B. Demonstrating consistent performance improvements based on a single seed run is fundamentally limited and risks being perceived as seed curation rather than robust generalization. Evaluating performance across multiple random seeds is necessary to establish statistical reliability.

  2. Out of the four datasets used in the evaluation, three are proprietary in-house datasets, which makes an objective verification of the reported contributions challenging. While the scarcity of public WSI datasets in cytopathology is acknowledged, releasing a subset or anonymized version of the private datasets should be considered to enable independent reproduction of the performance gains demonstrated by the proposed method.

  3. Table 1 appears to be a functional subset of Table 2. While I acknowledge that two additional FEs are evaluated in Table 1, the same information could be conveyed much more densely. For instance, the manuscript could either merge the unique rows into Table 2 or present the additional FE exploration in a smaller, compact table. Reclaiming this space would allow the manuscript to include more essential analytical experiments to strengthen the propose method.


Citations

[1] Chen, Richard J., et al. "Scaling vision transformers to gigapixel images via hierarchical self-supervised learning." 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2022.

[2] Deng, Ruining, et al. "Cross-scale multi-instance learning for pathological image diagnosis." Medical image analysis 94 (2024): 103124.

[3] Shin, Yonghan, and Won-Ki Jeong. "LRMIL: Efficient Low-Resolution Multiple Instance Learning via High-Resolution Knowledge Distillation for Whole Slide Image Classification." arXiv preprint arXiv:2606.06864 (2026). (MICCAI 2026)

[4] Hou, Wentai, et al. "H^ 2-MIL: exploring hierarchical representation with heterogeneous multiple instance learning for whole slide image analysis." Proceedings of the AAAI conference on artificial intelligence. Vol. 36. No. 1. 2022.

[5] Hong, Sungrae, et al. "Diagnose Like A REAL Pathologist: An Uncertainty-Focused Approach for Trustworthy Multi-Resolution Multiple Instance Learning." 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026.

[6] Xiong, Conghao, et al. "Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification." arXiv preprint arXiv:2301.08125 (2023).

[7] Li, Jiayun, et al. "A multi-resolution model for histopathology image classification and localization with multiple instance learning." Computers in biology and medicine 131 (2021): 104253.

[8] Pei Liu, Luping Ji, Xinyu Zhang, and Feng Ye. Pseudo-bag mixup augmentation for multiple instance learning-based whole slide image classification. IEEE Transactions on Medical Imaging, 43(5):1841–1852, 2024.

[9] Jiawei Yang, Hanbo Chen, Yu Zhao, Fan Yang, Yao Zhang, Lei He, and Jianhua Yao. Remix: A general and efficient framework for multiple instance learning based whole slide image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 35–45. Springer, 2022

[10] Gou, Jiaxiang, et al. "Queryable prototype multiple instance learning with vision-language models for incremental whole slide image classification." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 3. 2025.

[11] Ilse, Maximilian, Jakub Tomczak, and Max Welling. "Attention-based deep multiple instance learning." International conference on machine learning. Pmlr, 2018.

[12] Li, Bin, Yin Li, and Kevin W. Eliceiri. "Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning." 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2021.

[13] Wong, Bryan, et al. "Few-shot learning from gigapixel images via hierarchical vision-language alignment and modeling." Advances in Neural Information Processing Systems 38 (2026): 71313-71354.

[14] Shi, Jiangbo, et al. "Vila-mil: Dual-scale vision-language multiple instance learning for whole slide image classification." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

[15] Li, Jiawen, et al. "Dynamic graph representation with knowledge-aware attention for histopathology whole slide image analysis." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

[16] Huang, Peixiang, et al. "HyperPath: Knowledge-guided hyperbolic semantic hierarchy modeling for wsi analysis." International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2025.

Questions for authors

Please refer to the Critical Weaknesses section. I look forward to response addressing these points.

6 comments
AuthorsOpenReview ↗

Response to Reviewer NzUB

We sincerely thank the reviewer for the careful and knowledgeable review. We value the recognition that recovering intra-tile information from dense tokens adds no encoder cost, that cytopathology is an under-served domain, and that DI-MIL is a plug-and-play module across ViT-based encoders. The reviewer raised four substantive concerns: (1) positioning against hierarchical/multi-resolution and clustering-based MIL, together with motivation-matched baselines; (2) whether our critique of single-attention aggregation over-generalizes; (3) single-seed evaluation; and (4) reproducibility and presentation. We found each of them well-founded. We address them below with new experiments, and we commit to the corresponding revisions.

Protocol for all new results. Unless stated otherwise, every method uses the same Smart-CCS features, the same cross-validation folds (10-fold for Cervical, Serous Fluid and Respiratory; 5-fold for Oral2026), and the same training protocol as the paper. For each seed we average the test metric over folds. We then report the mean ± standard deviation across seeds. The comparisons against baselines (Table 2) use three seeds (1024, 2024, 3407), and every method has all three seeds. The Tile-versus-DI comparison (Table 1) uses five seeds (1024, 2024, 3407, 42, 7). All seeds that were run are reported; none were excluded.

AuthorsOpenReview ↗

Response to Reviewer NzUB - W1-A

W1-A. Hierarchical and multi-resolution approaches

"HIPT [1] progressively refines representations … class tokens across resolutions receive complex, contextualized attention scores via self-attention. … numerous studies have leveraged multi-resolution strategies to mitigate the oversmoothing of information [2-7]."

We agree that hierarchical and multi-resolution MIL are the closest related line of work and that the manuscript should position DI-MIL against it explicitly. We first clarify the conceptual relationship, then give a direct comparison.

Conceptual relationship. Multi-resolution methods [2–7] and DI-MIL both aim to expose evidence that a single tile embedding dilutes. They differ in where the additional granularity comes from:

  1. Multi-resolution methods add granularity before encoding. The WSI is re-tiled and re-encoded at several magnifications (or a high-resolution teacher is distilled into a low-resolution student [3]). Fine-scale evidence therefore requires additional image crops and encoder passes, and the fine-scale encoder sees a smaller field of view.
  2. DI-MIL adds granularity after encoding. Each tile is encoded once at its native size. Its dense tokens, each of which has already attended to the whole tile, are grouped into kk instances. Granularity increases without re-tiling, and every instance keeps the full tile context.
  3. The two are orthogonal. DI can be applied at any magnification of a multi-resolution pipeline: for example, by decomposing the dense tokens of each scale's tiles before cross-scale fusion.

Regarding HIPT [1] and single attention scores. We thank the reviewer for this point and agree that our statement should be made precise. The statement that each tile "can only receive a single weight" refers to the standard two-stage pipeline (one embedding per tile, then slide-level aggregation). In HIPT, the [CLS] tokens of 256-px patches are indeed contextualized by self-attention in the 4096-px region transformer. At each level of the hierarchy, however, a 256-px patch still enters the next level as a single [CLS] token, so its internal tokens are summarized before slide-level weighting. DI is complementary: it can decompose the 256-px ViT's dense tokens instead of forwarding only its [CLS] token. We did not include HIPT as a baseline. Its released weights come from hierarchical self-supervised pretraining on H&E 256/4096-px regions, and a faithful cytology counterpart would require re-pretraining both stages on cytology slides, which is beyond the scope of the rebuttal. We will make the scope of this statement explicit in Sections 1 and 3 and discuss HIPT in the related work.

Direct comparison with a multi-resolution MIL baseline. We implemented CS-MIL [2] with its official model and training recipe (cross-scale attention, cluster-stratified sub-bag sampling, gated-attention MIL). Its input uses Smart-CCS features at three concentric scales (300/600/1200 px, matching the official 2× ratio) on the same folds. We ran it on Respiratory and Oral2026, the two datasets on which our full tile-size ablation (Table 5) was already available:

AUC (3 seeds)RespiratoryOral2026Extraction cost relative to DI
Tile-ACMIL (1200 / 2048 px)80.17 ± 0.7087.45 ± 0.651×
CS-MIL [2] (300 + 600 + 1200 px)80.83 ± 0.7784.51 ± 1.06≈25× GPU-hours (Resp.) / ≈57× tiles (Oral)
DI-ACMIL (ours)88.39 ± 0.1987.89 ± 1.121× (+4% for clustering)

DI-ACMIL outperforms CS-MIL by +7.6 (Respiratory) and +3.4 (Oral2026) AUC points, and does so in every seed (6/6). CS-MIL needs features at all three scales. On Respiratory, extracting the 600-px and 300-px features costs 3.91× and 20.15× the GPU-hours of the native 1200-px features (Appendix A.9), so CS-MIL requires ≈25× the extraction compute of DI. On Oral2026 it requires ≈57× as many tile crops (Figure 2). This is consistent with Table 5, where single-scale finer tiling also stays below DI on both datasets while multiplying the extraction cost. We did not repeat multi-scale extraction on Cervical and Serous Fluid (≈1,900 and ≈1,600 tiles per slide), where the same 20–25× multiplier would apply. Avoiding exactly this cost is the practical motivation for DI.

Revision. We will add a dedicated related-work paragraph on hierarchical and multi-resolution MIL covering [1–7]. We will state the before-encoding versus after-encoding distinction and the complementarity above, and we will add CS-MIL to Table 2 and Table 5.

AuthorsOpenReview ↗

Response to Reviewer NzUB - W1-B

W1-B. Clustering-based MIL and motivation-matched baselines

"Multiple studies have already incorporated clustering for instance selection and aggregation within MIL [8-10]. … the clear distinction … remains insufficiently articulated. … such clustering-based approaches should be included as direct baselines."

We agree that the distinction must be stated precisely.

Distinction. In short, clustering-based MIL methods reorganize the bag, whereas DI-MIL redefines the instance.

PseMix [8]ReMix [9]QPMIL-VL [10]PTCMIL (Table 2)DI-MIL
What is clusteredtile embeddings across a slidetile embeddings across a slidetile embeddings vs. learned prototypestile embeddings across a slidedense tokens within one tile
Purposepseudo-bag mixup augmentationbag reduction and latent augmentationprototype memory for incremental learningprompt-guided slide prototypesfiner instances
Atomic unit seen by the aggregatortile (or coarser)prototype (coarser than tile)prototypeprototypesub-tile instance
Information mixed inside a tile embeddingnot recoverablenot recoverablenot recoverablenot recoverableseparated before pooling

All four methods operate on tile embeddings that have already been pooled. Once a sparse abnormal cell has been averaged with the surrounding background into one tile vector, no subsequent clustering across tiles can separate it again. DI-MIL operates one step earlier, inside the tile, on the dense tokens that the frozen encoder already computes. It therefore changes what enters the bag, while clustering methods change how the bag is organized or augmented. The two can be combined: ReMix, for example, can build its prototypes from DI instances.

Direct baselines. Table 2 of the submission already includes PTCMIL. We have now added ReMix [9] with its official code (per-slide K-means prototypes, covariance-based latent augmentation, official ABMIL head):

AUC (3 seeds)CervicalSerous FluidRespiratoryOral2026
PTCMIL94.63 ± 0.0284.43 ± 0.4979.36 ± 0.4986.10 ± 1.47
ReMix [9]93.43 ± 0.3283.67 ± 0.1480.95 ± 0.6084.23 ± 2.36
DI-ACMIL95.52 ± 0.1788.12 ± 0.1588.39 ± 0.1987.89 ± 1.12

DI-ACMIL outperforms both clustering-based methods on all four datasets and in every seed (12/12 dataset–seed pairs against each method). We did not run QPMIL-VL [10]: it requires a pathology vision-language encoder with class-level text prompts and targets class-incremental learning, which differs from the frozen-vision-encoder, fixed-label setting studied here. We will discuss it, together with PseMix [8], in the related work.

Motivation-matched baselines. We agree that DTFD-MIL and MambaMIL address different problems. Including the additions in this response, Table 2 now contains, for each family relevant to DI-MIL's motivation:

  • multi-resolution: CS-MIL [2];
  • clustering and prototypes: ReMix [9], PTCMIL;
  • graph-based relational modeling: WiKG [15] (newly added, discussed in W2);
  • the original attention, sequence and pseudo-bag baselines.

Revision. We will add ReMix, CS-MIL and WiKG to Table 2, insert the comparison table above into the related work, and state the "reorganize the bag versus redefine the instance" distinction explicitly in Sections 1 and 2.

Reviewer_xZMtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Different from conventional WSI multiple-instance learning uses the encoder tile as the basic unit of slide-level aggregation, this paper proposes Decomposed-Instance MIL (DI-MIL) to implement smaller tiles and the original tile extraction and encoder field of view. Experiments on cytopathology benchmarks show that decoupling encoding context from instance granularity outperforms both standard MIL baselines and naive extraction of smaller source tiles.

Critical strengths
  • This plug-and-play DI-MIL module is applicable to various patch-level foundation models.
  • The method operates at a finer instance granularity and enables richer knowledge modeling, avoiding the need to re-encode the WSI using smaller patch sizes.
Critical weaknesses
  • This paper appears to present a general-purpose solution for MIL problems. Can it be extended to non-medical fields?
Questions for authors
  • This method actually denotes a deeper decomposition of WSI tiles. Does the decomposition stragety show biological significance? I think it will be helpful if there is more computational analysis.

  • How does this proposed method differ from aggregation modeling in other scales, such as single-cell level?

  • This idea seems generalizable to MIL datasets in other fields. Please provide additional experiments and analysis.

2 comments
AuthorsOpenReview ↗

Response to Reviewer xZMt

We thank the reviewer for the positive assessment. All new results use the same protocol for Tile and DI and report mean ± std over seeds.

Q1: This method actually denotes a deeper decomposition of WSI tiles. Does the decomposition stragety show biological significance? I think it will be helpful if there is more computational analysis.

We provide evidence at three levels.

  1. Dense tokens retain cell-level evidence (paper, A.2). On 18 public cytology datasets with cell boxes (35,014 cells), annotated cells are recovered from frozen dense tokens well above random. Cytopathology backbones do better than DINOv3: insertion gain 0.225 vs 0.200, with Smart-CCS ahead on 14/18 datasets (p = 0.0014). This matches the larger DI gains we see with these backbones.
  2. DI clusters follow cell boundaries (new). We applied DI's k-means (k = 4) to crops centred on annotated cells. The best-matching clusters cover the cell core with IoU 0.283, vs 0.149 with random cluster labels and 0.247 against another cell's box. The observed clusters win on 8/8 datasets. The grouping criterion also matters downstream: on Respiratory (5 seeds), k-means DI reaches 88.2 AUC, vs 86.3 for spatial 2×2, 84.0 for random grouping, and 80.5 for Tile.
  3. Trained DI-MIL attends to abnormal cells (new). We trained on an extra cytology cohort (named Cervical-2, 944 WSIs) with slide labels only (5-seed), then scored attention against expert abnormal-cell annotations on 63 positive WSIs within the cohort. Abnormal cells cover only 0.013% of the dense patch tokens. Lift is the enrichment of abnormal cells among the top-attended B% of slide area; chance is 1×.
ACMILTileDIDI better (slides)p
Instance AUROC0.460.5656/631e-10
Lift@1%0.26×8.5×51/634e-10
Lift@10%0.63×2.2×60/631e-11

Slide-level AUC is the nearly the same for both (tile 95.4 vs DI 95.6), so the gap reflects where the model looks. Tile attention is at chance, while DI isolates abnormal cells. With ABMIL the two are comparable at the top 1% (5.3× vs 4.8×), and Tile is better at broader budgets. Decomposition makes cellular evidence addressable, and ACMIL's multi-branch attention exploits it best. These results will supplement the qualitative-only A.5.

Q2: How does this proposed method differ from aggregation modeling in other scales, such as single-cell level?

  • No detector, annotations, or extra encoding. Single-cell MIL needs a cell detector, usually trained on cell labels, and one encoder pass per cell crop. Detection errors propagate, and crops lose their surroundings. DIs are formed after encoding from tokens that have attended to the whole tile, so cytological context such as clusters and background is kept. Extra extraction cost is 1.04× (A.9).
  • Adaptive granularity. A DI is a set of feature-similar tokens. It can be one abnormal cell, a cell cluster, or background, i.e. between tile and cell scale, as Q1(2) shows.
  • The cell-scale analogue is costly. At 0.5 µm/px, each Smart-CCS token spans about 37.5 µm, roughly one cell. Using every token as an instance means a 256N bag: 457 GiB vs 7.1 GiB of features on Oral2026 (A.8). Finer tiling toward cell scale costs 20× the GPU-hours at 300 px and still trails DI (85.5 vs 88.9 AUC on Respiratory, Table 5).

DI changes instance construction, not aggregation, so it is complementary to cell-level and multi-scale methods. We will add this to the related work.

Weakness: This paper appears to present a general-purpose solution for MIL problems. Can it be extended to non-medical fields?

Q3: This idea seems generalizable to MIL datasets in other fields. Please provide additional experiments and analysis.

We built weakly supervised MIL tasks on DOTA-v1.0 aerial images (1,869 images, 448-px tiles; an image is positive if it contains the class). We used generic DINOv3, 5-fold CV, 5 seeds, and no change to the method.

AUC (%)TileDISeeds won
Roundabout, ABMIL91.2 ± 0.992.6 ± 0.95/5
Roundabout, ACMIL91.6 ± 0.692.3 ± 0.45/5
Ship, ABMIL / ACMIL97.3 / 97.497.3 / 97.5tie

DI consistently improves roundabout, where a single object is the only evidence, and matches the near-ceiling baseline on ship. It transfers to a non-medical domain and a generic backbone without tuning, while its largest gains remain where evidence is sparse, as in cytopathology. This will be added to the appendix.

Thanks for the detailed response. My concerns have been addressed.

Reviewer_6e6hOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DI-MIL, a framework that decouples visual encoding context from instance granularity in weakly supervised whole-slide image classification. Instead of representing each image tile with a single global embedding, DI-MIL clusters dense spatial tokens extracted by a frozen foundation model into multiple groups using spherical k-means. Each cluster is aggregated into an independent instance embedding, allowing attention-based MIL aggregators to assign separate weights to different sub-tile representations.

The method is evaluated on four cytopathology datasets using three foundation-model backbones and two MIL aggregators. The results demonstrate generally improved classification performance compared with conventional tile-level representations. Ablations examine the number of decomposed instances, alternative grouping strategies, and direct smaller-tile extraction. Additional analyses investigate dense-token representation quality, statistical consistency, and computational costs.

Critical strengths
  1. The paper identifies a meaningful limitation of conventional MIL: the spatial context used for visual encoding need not determine the granularity of downstream instance-level attention. This distinction is particularly relevant in cytopathology, where sparse abnormal cells may be diluted within large image tiles. The proposed decomposition strategy provides a simple and practical way to exploit dense foundation-model representations without additional encoder forward passes or modifications to existing MIL aggregators.

  2. The experimental validation is relatively comprehensive and includes several important controlled comparisons. The authors evaluate different foundation models and MIL aggregators, compare feature-based grouping against random and spatial grouping, and distinguish the effects of dense-token representations from increasing instance granularity. The smaller-tiling comparisons and resource analyses further support the practical advantages of post-encoding decomposition.

Critical weaknesses
  1. The evidence that decomposed instances capture diagnostically relevant cellular information remains insufficient. Spherical k-means groups tokens according to feature similarity, which does not necessarily correspond to malignant cells or clinically meaningful structures. Although the paper includes qualitative attention visualizations and analyses of dense-token representation quality, it does not directly validate whether highly attended decomposed instances localize diagnostic cells. Quantitative evaluation against cell-level annotations would more directly support the central motivation.

  2. The methodological novelty and generalizability require stronger justification. DI-MIL primarily combines existing token clustering and feature pooling operations, and its strongest improvements are concentrated in cytopathology-specific backbones, while gains with DINOv3 are relatively modest. Furthermore, all four downstream benchmarks involve binary cytopathology classification. The current evidence therefore does not fully establish whether the approach offers a broadly applicable instance-construction principle beyond this setting.

Questions for authors
  1. Have the authors quantitatively evaluated whether high-attention decomposed instances correspond to malignant cells or diagnostically informative regions? Can existing cell-level annotations be used to measure localization precision or enrichment relative to conventional tile-level attention?

  2. What distinguishes DI-MIL fundamentally from existing feature-space token clustering or prototype-based MIL approaches beyond retaining multiple cluster embeddings per tile? Can the authors clarify which aspect of the proposed instance-construction formulation represents the principal methodological advance?

  3. Why are the improvements substantially larger with Smart-CCS than with DINOv3? Does this indicate that DI-MIL depends on domain-specific dense-token representations, and what conclusions can be drawn about its generalizability to other pathology foundation models?

  4. Are results available on conventional H&E histopathology datasets or multiclass WSI classification tasks? Such evidence would help establish whether the proposed instance decomposition is broadly useful beyond the four binary cytopathology benchmarks.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 6e6h - Q1

We thank the reviewer for the constructive comments. We address each question below. Unless otherwise stated, new results are reported as mean ± std over five random seeds (1024, 2024, 3407, 42, 7), each averaged over cross-validation folds.


Q1. Quantitative localization of high-attention decomposed instances

"Have the authors quantitatively evaluated whether high-attention decomposed instances correspond to malignant cells or diagnostically informative regions? Can existing cell-level annotations be used to measure localization precision or enrichment relative to conventional tile-level attention?"

Appendix A.2 evaluates, on 18 public cell-annotated datasets, whether frozen dense tokens recover annotated cells; this analysis operates at the encoder level. Following the reviewer's suggestion, we additionally evaluate localization at the level of trained MIL attention.

Protocol. We use an additional cervical cytology cohort (Cervical-2) that provides both slide-level diagnoses and expert bounding-box annotations of abnormal cells. ACMIL is trained on Smart-CCS features with slide-level labels only (944 slides, 5-fold cross-validation, five seeds), using Tile and DI bags under an identical protocol. Cell annotations are used exclusively for evaluation, on 63 positive test slides whose annotations are registered to the tile grid. A token is labeled positive if its center lies within an abnormal-cell box; positive tokens constitute only ~0.02% of all tokens per slide. Instances are ranked by attention, and we report:

  • Instance AUROC: the AUROC of attention scores for discriminating instances that contain abnormal-cell tokens.
  • Lift@B%: the abnormal-token precision of the highest-attention instances covering B% of the slide's tokens, normalized by the slide-level abnormal-token fraction (1 corresponds to random ranking). Budgets are area-matched, so both instance types are compared on equal tissue area.
InstanceInst. AUROCLift@1%Lift@5%Lift@10%
Tile0.4580.260.630.63
DI0.5628.533.172.17

The top 1% of attended DI instances are enriched 8.5-fold for abnormal cells, whereas tile-level attention performs at chance level (0.26-fold). DI achieves a higher instance AUROC on 56 of 63 slides (Wilcoxon signed-rank test, p = 1.4 × 10⁻¹⁰). Slide-level AUC is comparable between the two bags (95.43 vs. 95.60), indicating that the difference arises from the localization of attention rather than from classification performance.

We will add this analysis to the appendix and reference it in Section 4 alongside the qualitative attention maps of Appendix A.5.

AuthorsOpenReview ↗

Response to Reviewer 6e6h - Q2

Q2. Distinction from token-clustering and prototype-based MIL

"What distinguishes DI-MIL fundamentally from existing feature-space token clustering or prototype-based MIL approaches beyond retaining multiple cluster embeddings per tile? Can the authors clarify which aspect of the proposed instance-construction formulation represents the principal methodological advance?"

In short: existing clustering and prototype MIL methods reorganize the bag; DI-MIL redefines the instance. Prior methods cluster across tiles, on embeddings that have already been pooled, in order to compress or restructure a bag whose atomic unit is still the tile. DI-MIL clusters within each tile, on dense tokens, before aggregation, so that the atomic unit itself becomes finer than the tile. The two differ in three respects:

  1. Direction of granularity. Prototype and clustering methods (e.g., ReMix, PseMix, PTCMIL) map N tile embeddings to fewer, coarser units. DI-MIL maps N tiles to kN sub-tile instances. Retaining multiple cluster embeddings per tile is therefore not an implementation detail; it is what allows attention to act below the tile level, which no coarsening scheme can provide.
  2. Point of operation. Once a tile is pooled into a single embedding, evidence mixed with background inside that tile cannot be recovered by any subsequent clustering. DI-MIL operates before this information loss, on the dense tokens that frozen encoders already compute.
  3. Role in the pipeline. DI-MIL determines what enters the bag, whereas clustering and prototype methods determine how the bag is aggregated. The two are therefore orthogonal.

Principal advance. We identify that conventional MIL couples two distinct design choices: the context seen by the encoder and the granularity at which evidence is weighted. DI-MIL decouples them. The source tile remains the unit of encoding, while the unit of weighting is derived from its dense tokens after encoding, at no additional encoder cost and without modifying the aggregator. Intra-tile clustering is the means of realizing this decoupling, not the contribution in itself.

Empirical support. We further compared DI-MIL with ReMix, a representative clustering-based MIL method, under identical Smart-CCS features and splits (five seeds):

AUCCervicalSerous FluidRespiratoryOral2026
ReMix93.48 ± 0.3783.82 ± 0.2481.16 ± 0.5284.00 ± 1.88
DI-ACMIL95.58 ± 0.1788.09 ± 0.4388.22 ± 0.3188.31 ± 1.05

DI-ACMIL also outperforms PTCMIL, which clusters prompt-guided tile representations at the slide level, in AUC on all four datasets and all three backbones (Table 2). The localization results in Q1 confirm the underlying mechanism: with DI, attention concentrates on abnormal cells (8.5-fold enrichment), whereas tile-level attention does not.

We will revise the related-work section and the introduction to state this distinction and the decoupling principle explicitly, and include ReMix as an additional baseline.

AuthorsOpenReview ↗

Response to Reviewer 6e6h - Q3

Q3. Backbone dependence of the gains

"Why are the improvements substantially larger with Smart-CCS than with DINOv3? Does this indicate that DI-MIL depends on domain-specific dense-token representations, and what conclusions can be drawn about its generalizability to other pathology foundation models?"

DI-MIL can separate only the evidence that the backbone's dense tokens already distinguish. Its gain therefore reflects the quality of the dense representation for the relevant morphology, rather than a dependence on a particular domain. Appendix A.2 analyzes this directly:

  • Quantitative. Across 18 public cell-annotated datasets, annotated cells are most readily recovered from Smart-CCS tokens (insertion gain G = 0.225), followed by UniCAS (0.217) and DINOv3 (0.200). Smart-CCS and UniCAS outperform DINOv3 on 14/18 and 15/18 datasets, respectively (one-sided paired Wilcoxon p = 0.0014 and 0.00042).
  • Qualitative. Figure 4 shows that the cytopathology-specific backbones assign distinct token representations to abnormal cells. With DINOv3, abnormal and normal cells are less separable in token space, so intra-tile clustering more frequently assigns them to the same group, and the resulting instance remains a mixture.

The ranking of DI-MIL gains coincides with this ranking. In response to R3, we also conduct the five-seed results, which further show that the gain under DINOv3, although smaller, is consistently positive:

BackboneMean ΔAUC (DI − Tile)DI better (seeds × datasets × aggregators)Token-level evidence recovery G (A.2)
Smart-CCS+2.8032/400.225
UniCAS+1.6837/400.217
DINOv3+0.6932/400.200

We draw three conclusions:

  1. DI-MIL requires neither retraining nor a specific encoder, and it improves all three backbones, including a general-domain model.
  2. The magnitude of the improvement scales with the ability of the dense tokens to resolve diagnostic structures. This is a property of the representation rather than a limitation of the formulation, and DI-MIL is expected to benefit from foundation models with stronger dense representations.
  3. The improvement is not specific to a single foundation model: it holds for two different cytopathology foundation models. Results with an H&E pathology foundation model are part of the experiments described in Q4.

We will incorporate this analysis into Section 4.2 and state explicitly that the expected gain scales with dense-token quality.

Comments on the paper

1 comment
Area_Chair_G84sOpenReview ↗

A Gentle Nudge: Let's Keep the Discussion Going!

Dear Reviewers,

The authors have now posted their responses, addressing your concerns and incorporating your suggestions. We are already halfway through the discussion period!

One of the distinguishing features of ICLR is its interactive review process, where constructive discussions not only help strengthen individual papers but also contribute valuable insights to the broader scientific community.

So, let's not let the discussion be a one-way street! I encourage you to revisit the authors' responses, acknowledge their efforts, and engage in further discussion where needed. After all, a good scientific discussion is often as valuable as the final verdict!

-AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.