List of Papers Browse by Subject Areas Author List
Abstract
The application of large vision-language models to computational pathology holds great promise for diagnostic assistants but faces a critical computational bottleneck: the gigapixel scale of Whole Slide Images (WSIs). A single WSI typically contains over 10^5 patches, creating sequence lengths that exceed the constraints of standard Transformer architectures. Existing solutions often resort to spatial sampling, which risks discarding diagnostically critical evidence. To address this, we propose TC-SSA (Token Compression via Semantic Slot Aggregation), a learnable token compression framework that aggregates patch features into a fixed number of semantic slots. A gated routing module assigns patches to slots using sparse Top-2 routing, followed by weighted aggregation, enabling global slide coverage under a strict token budget. The resulting representation retains diagnostically relevant information while reducing the number of visual tokens to 1.7% of the original sequence. On SlideBench (TCGA), our model achieves 78.34% overall accuracy and 77.14% on the diagnosis subset, outperforming sampling-based baselines under comparable token budgets. The method also generalizes to MIL classification, reaching AUCs of 95.83% on TCGA-BRCA and 98.27% on TCGA-NSCLC, and an ISUP grading accuracy of 79.80% on PANDA. These results suggest that learnable semantic aggregation provides an effective tradeoff between efficiency and diagnostic performance for gigapixel pathology reasoning. Our code is available at https://github.com/OzzyChen97/TC-SSA.git.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2663_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/OzzyChen97/TC-SSA.git
Link to the Dataset(s)
SlideChat dataset: https://huggingface.co/datasets/General-Medical-AI/SlideChat/blob/main/README.md
TCGA dataset: https://www.cancer.gov/ccg/research/genome-sequencing/tcga
BCNB dataset: https://bcnb.grand-challenge.org
BibTex
@InProceedings{CheZhu_TCSSA_MICCAI2026,
author = { Chen, Zhuo AND Yang, Xiaoyu AND Xu, Lijian},
title = { { TC-SSA: Token Compression via Semantic Slot Aggregation for Gigapixel Pathology Reasoning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This manuscript focused on the scalability challenge caused by the gigapixel scale of WSIs when applying the Transformer architecture to downstream tasks in computational pathology. To face this challenge, the authors proposed Token Compression via Semantic Slot Aggregation (TC-SSA), a learnable token-budgeting framework aggregating all patch features into fixed learnable semantic slots. The main contributions of the proposed TC-SSA are summarized as follows:
- Semantic Slot-Based Token Compression. Based on the MoE-like routing technique, this module is proposed to route all visual tokens into a fixed number of semantic slots for feature aggregation, while preserving meaningful contextual information in WSIs.
- Robust Regularization for Semantic Slots. To guarantee the stability of the slot distribution from the above module. The authors proposed a robust regularization loss to mitigate slot collapse and ensure routing stability during training. Comprehensive experiments are done across diverse benchmarks and public datasets to prove the effectiveness of the proposed modules in TC-SSA. Results proved that, under the compressed tokens derived from TC-SSA, the models achieved a trade-off between efficiency and performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- This manuscript is well organized and follows a clear logical flow. Overall, the contribution of this work is satisfying.
- The authors have provided an anonymous link to open-source code, demonstrating the reliability and reproducibility of the reported results in this work.
- The motivation of TC-SSA is quite meaningful to the facing the challenge of gigapixel WSIs when applying large-scale vision and language model in computational pathology.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
The boundary of novelty could be clarified more clearly. In this work, several design choices, such as semantic slots and Top-2 routing, appear closely related to techniques commonly used in Mixture-of-Experts in the training of foundation model. It would be helpful for the authors to better clarify how the proposed design specifically adapts in token compression for WSIs.
-
The conceptualization of the “semantic affinity clustering objective” is somewhat confusing. Based on the current formulation, it is difficult to see the relationship with a clustering process. It appears more like a loss design for improving slot balance and routing stability. A more precise conceptualization is needed for the demonstrations of method. Additionally, the loss in equation (2) and (3) are similar with the loss design in MoE, the difference should be emphasized.
-
A more detailed analysis of this hyperparameter auxiliary loss weight λ would help readers better understand the stability and robustness of the proposed loss function (although the authors have provided the conclusion about what will happen when reducing λ).
-
The visualization of Figure 3.It is currently difficult to understand what exactly is being visualized and how this figure supports the conclusion of meaningful semantic slots related to WSIs.
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I think this manuscript solves an interesting and practical problem in large visual-language models for computational pathology. The major factors in the current form of this manuscript are the novelty boundary (looks like an application of MoE-based techniques?), and missing validations for some key parameters. I suggest that the authors could provide more elaboration on the concerns in the Weaknesses section, which would help further strengthen my confidence in the manuscript.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Most of my concerns are resolved during rebuttal, and the boundary of novelty is clearly claimed with more elaborations on the result details. I will accept this manuscript for publication.
Review #2
- Please describe the contribution of the paper
To address the computational bottleneck in WSI-level tasks, this manuscript proposes TC-SSA, a novel learnable token compression framework that aggregates patch features into a fixed number of semantic slots. Specifically, a gated routing module is utilized to route patches to slots, followed by weighted aggregation for downstream WSI-level tasks. Extensive experiments on the TCGA benchmark demonstrate that TC-SSA effectively improves the performance of VQA and classification tasks while maintaining low computational FLOPs.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.This manuscript proposes TC-SSA, which leverages learnable semantic slots for efficient patch feature compression, effectively tackling the critical computational bottleneck in gigapixel WSI-level tasks. 2.This manuscript introduces a set of robust regularization terms to mitigate slot collapse and ensure routing stability. 3.This manuscript provides extensive experimental validation on downstream multiple-choice VQA and classification tasks, demonstrating that TC-SSA substantially reduces computational overhead while boosting performance.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The manuscript lacks individual ablation experiments for loss terms. 2.Building on the first weakness, the reviewer is particularly concerned about the validity of the load-balancing loss in this specific domain. In a single WSI, the number of background patches and diagnostically critical patches varies drastically (i.e., extreme inherent imbalance among patches with different semantics). Under such circumstances, is it biologically and methodologically sound to force a uniform distribution across slots? The reviewer strongly suggests an independent ablation study on the load-balancing loss to clarify its actual impact. 3.Figure 3 lacks sufficient explanation. The three different categories of patches do not appear to be well-separated in the visualization based on t-SNE. Furthermore, it is unclear whether the three displayed subplots represent the t-SNE results within three distinct semantic slots. 4.TC-SSA completely discards the learning of the local microenvironment and overall structures. Would this loss of spatial geometry lead to severe performance degradation in some downstream tasks, such as survival prediction? 5.In Section 3 (Ablation Studies), the authors explicitly claim to have conducted experiments replacing the CONCH encoder with the UNI encoder. However, the corresponding numerical results are entirely missing from the manuscript.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This manuscript proposes TC-SSA, a novel learnable token compression framework that aggregates patch features into a fixed number of semantic slots. Although empirical results demonstrate that TC-SSA achieves significant performance improvements while maintaining low computational FLOPs, the manuscript lacks rigorous ablation studies for its key contributions (the regularization terms). Furthermore, the inadequate explanation of visualizations (Weakness 3) and the absence of results to support some claims (Weakness 5) undermine the overall credibility of the manuscript. If the authors can address the weaknesses in the rebuttal, the reviewer will consider raising the final score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The reviewer thanks the authors for their response and explanations. Weaknesses 1, 2, 4, and 5 have been satisfactorily addressed. However, the clarification regarding Weakness 3 remains inadequate. Specifically, in each subplot of Figure 3 (representing different TCGA cohorts), patches assigned to a specific semantic slot do not exhibit distinct grouping patterns after t-SNE. Thus, it fails to effectively validate the claim that patches routed to the same slot cluster together in the embedding space.
This manuscript proposes TC-SSA, a novel learnable token compression framework that aggregates patch features into a fixed number of semantic slots. Empirical results demonstrate that TC-SSA achieves substantial performance improvements while maintaining low computational FLOPs. The technical contributions of this work are solid enough for acceptance at MICCAI 2026.Nevertheless, the manuscript requires further refinement in its writing to more clearly analyze and present results.
Review #3
- Please describe the contribution of the paper
Standard VLMs cannot process full WSIs due to quadratic attention costs. Sampling-based approaches discard most patches. TC-SSA offers a middle ground, global coverage under strict token constraints, enabling practical deployment of VLMs on gigapixel pathology images.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Novel formulation of token compression as semantic routing
Rather than spatial sampling, patches are routed to learned semantic slots based on content affinity. This aggregates diagnostically relevant regions scattered across the slide; a meaningful departure from prior sampling or attention-pooling approaches.
2.Extreme compression with competitive performance
Achieving 78.34% accuracy with only 32 tokens (58× compression) is strong. TC-SSA surpasses the uncompressed SlideChat on Diagnosis (77.14% vs 73.27%), suggesting compression can remove noise and improve signal for reasoning tasks.
3.Principled regularization for routing stability
The auxiliary loss addressing slot collapse is well-motivated and empirically validated; performance degrades when removed, confirming it’s necessary rather than added complexity.
4.Consistent generalization
Strong results across VQA benchmarks, MIL classification, different encoders (CONCH, UNI), and zero-shot transfer suggest the approach is robust rather than benchmark-specific.
5.Practical efficiency
O(N·K) complexity with K=32 makes deployment tractable under real clinical memory and latency constraints
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Limited comparison with learnable compression methods
Baselines are sampling-based (LLaVA-Med, Quilt-LLaVA) or uncompressed (SlideChat). Comparison with other learnable approaches; Perceiver-style cross-attention, Q-Former, or cited methods like LOC-PATH [7] and WSISum [20]; is missing.
2.Unexplained Microscopy performance gap
TC-SSA underperforms SlideChat on Microscopy (81.94% vs 87.64%) but this isn’t analyzed. Why does semantic aggregation lose fine-grained morphological information?
3.Lack of qualitative analysis
No visualization of patch-to-slot routing, failure cases, or slot specialization. The t-SNE (Figure 3) is too coarse to validate semantic grouping claims.
4.Arbitrary hyperparameter selection
K=32 is chosen empirically with no guidance on scaling with slide complexity. The K=128 degradation (“over-fragmentation”) isn’t explained.
5.Incomplete efficiency evaluation
Only FLOPs reported, no inference latency or memory usage despite clinical deployment claims.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important problem with a well-motivated solution. Semantic slot routing is a reasonable alternative to spatial sampling, and the results are solid; 58× compression while outperforming the uncompressed baseline on diagnostic tasks.
Supporting factors:
Novel semantic routing formulation Strong compression-accuracy tradeoff Consistent generalization across benchmarks and encoders Limiting factors:
Missing comparisons with other learnable compression methods (Perceiver, Q-Former, LOC-PATH) Unexplained Microscopy performance gap No qualitative slot analysis or failure cases
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the ACs and reviewers for constructive feedback. We are deeply encouraged by your recognition, especially: 1) the novel semantic routing design for WSI token compression; 2) strong efficiency with high compression and competitive performance; 3) effective regularization for stable routing; and 4) extensive experiments and reproducible results.
And, we have provided detailed clarifications and additional context to the primary concerns:
1.Figure 3 Explanation (R1W4, R2W3, R3W3). Each subplot corresponds to a different TCGA cohort. Each point is a patch embedding projected via t-SNE, with each color representing one semantic slot. The visualization shows patches routed to the same slot form coherent clusters across datasets, suggesting the gate consistently groups similar tissue patterns (tumor, stroma, necrosis).
2.Missing Results and Baselines (R2W5, R3W1). We clarify Table 2 uses UNI encoder as our ablation study. We now directly compare table 2 TCGA-BRCA with WSISum[20], which uses CONCH. We report comparisons using its cited baseline: ABMIL Full WSI scores 0.858 vs. WSISum 0.874; TransMIL Full WSI scores 0.843 vs. WSISum 0.818.Which confirm that our baselines are stronger representatives and that the cited methods do not outperform standard pooling on these benchmarks. Besides, we result LOC-PATH[7] achieving only 51.9% accuracy on SlideBench zero-shot, far below TC-SSA’s 78.34%.
3.Difference with MoE (R1W1). MoE routes tokens to expert networks, retaining the original sequence length with only Top-K experts active per token. TC-SSA routes patches to semantic slots for information compression (K«N) via weighted pooling(Eq. 1). Critically, all K slots are active outputs that participate in downstream reasoning, unlike MoE’s partial expert activation. TC-SSA borrows the routing intuition from MoE but repurposes it for semantic consolidation, not selective computation.
4.Semantic Affinity Clustering and Load-Balancing (R1W2, R2W1,2). We do not perform explicit clustering, patches with similar semantics naturally aggregate into the same slot as an emergent behavior of routing. L_switch is a soft regularizer that only penalizes the degenerate case where patches flood into one slot, it does not force equal distribution. The gate can still assign background patches to few slots while diagnostic regions occupy others. Without L_switch, the gate collapses: reducing λ causes the last 5 slots to receive nearly no patches. We will revise terminology to semantic slot regularization.
5.Loss Ablation and λ (R1W3, R2W4). Many thanks for this suggestion. We provide additional analysis: λ=0 yields 76.29% accuracy, while λ=0.5 yields 78.75%, confirming that increasing λ improves performance. We argue that when λ is too small, patches concentrate in only a few slots, under-utilizing the full semantic capacity. Entropy weight 0.5 and α=10^-4 follow standard sparse routing practice for numerical stability.
6.Spatial Geometry and Microscopy Gap (R2W4, R3W2). Spatial information is not entirely discarded. Each patch is extracted from a specific spatial location, so its embedding already captures local context. The impact of compression is task-dependent. Microscopy relies on fine-grained morphological detail, where compression incurs a 5.7% gap and outperforms all other compressed methods (vs. MedDr: 73.30%). Diagnosis relies on overall tissue composition, where TC-SSA (77.14%) surpasses the uncompressed SlideChat (73.27%).
7.K and Efficiency (R3W4,5). After extracting the local and global features of WSIs, the average response time was within 1 second, and GPU memory consumption was approximately 15GB(Slidechat 27GB). K=32 is validated by ablation (Figure 4): stable for K ∈ {16, 32, 64}, with degradation at K=128 due to over-fragmentation. TC-SSA’s O(NK) complexity guarantees tractable clinical inference.
Thanks again for your effort! If you find our response satisfactory, could you please consider improving our score?
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Please try to address the pointed weaknesses during the rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers appreciate the motivation and extensive experimental results, and the initial concerns have been fully addressed during the rebuttal. I recommend accepting this submission.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes a semantic slot aggregation framework for token compression in gigapixel pathology reasoning. Reviewers appreciate the practical importance of the problem. The rebuttal clarifies most concerns regarding the novelty boundary, regularization design, missing result descriptions, and the trade-off between compression and fine-grained reasoning. Some concerns remain about the clarity of slot visualization and presentation, which should be addressed in revision.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes TC-SSA, a novel learnable token compression framework that aggregates patch features into a fixed number of semantic slots. The reviewers generally recognize the technical merit of the proposed method, and the empirical results demonstrate a favorable balance between predictive performance and computational efficiency. Most of the concerns raised during the review process have been satisfactorily addressed through the rebuttal.
One remaining concern relates to the interpretation of the t-SNE visualization in Figure 3.The current evidence does not fully support the claim that patches assigned to the same semantic slot form distinct clusters in the embedding space, and this aspect would benefit from further analysis and clarification. Nevertheless, this issue does not substantially undermine the overall contribution of the work.
Overall, the proposed framework is technically sound, well-motivated, and demonstrates meaningful practical value for efficient WSI analysis. While the manuscript would benefit from clearer presentation and discussion of the results, I believe the paper meets the acceptance bar for MICCAI and support its acceptance.