List of Papers Browse by Subject Areas Author List
Abstract
In computational pathology (CPath), developing omni-modal self-supervised learning (SSL) models that integrate histology, genomics, and clinical reports enables transferable representation learning for whole slide images (WSIs). Existing approaches implicitly force heterogeneous modalities into a uniform latent space by contrastive alignment, causing modality collapse where unique, synergistic diagnostic signals (termed as Φ) are discarded in favor of trivial redundancy. We hypothesize that the strongest task-agnostic SSL training signal stems from distilling the synergistic interactions over merely aligning shared redundancy. To this end, we introduce Φ-OMNI, a synergistic information disentanglement framework grounded in Partial Information Decomposition (PID) theory for slide representation learning. Unlike standard contrastive approaches, Φ-OMNI employs a Synergistic Information Bottleneck (SIB) regulated by the proposed ΦID objective, which explicitly suppresses marginal redundancy while maximizing irreducible synergy, thereby distilling high-order cross-modal interactions. Following pretraining on breast (n=1031) and lung (n=919) cohorts, Φ-OMNI demonstrates superior few-shot performance across five independent external datasets spanning eight tasks compared to supervised and SSL baselines. Source code is available.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0048_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/imicjs/PhiOmni
Link to the Dataset(s)
TCGA dataset: https://portal.gdc.cancer.gov/
CPTAC dataset: https://dctd.cancer.gov/research/networks/cptac
BRACS dataset: https://www.bracs.icar.cnr.it/
BibTex
@InProceedings{LiuMin_Synergistic_MICCAI2026,
author = { Liu, Mingxin AND Cai, Chengfei AND Lu, Anwen AND Xu, Pengbo AND Li, Jun AND Li, Jinze AND Chen, Depin AND Xu, Jun},
title = { { Synergistic Information Disentanglement for Omni-Modal Slide Representation Learning in Computational Pathology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper addresses an important limitation in multimodal pathology pretraining: existing methods often rely on cross-modal alignment in a way that may wash out complementary, modality-specific information. To tackle this, the authors propose \Phi-Omni, an omni-modal self-supervised framework that uses histology, transcriptomics, and pathology reports during pretraining, but only needs WSI input at test time. The main idea is to replace redundancy-driven alignment with a synergy-oriented objective, implemented through a Synergistic Information Bottleneck (SIB) and a \PhiID loss inspired by Partial Information Decomposition. The method is evaluated on breast and lung cohorts and obtains strong few-shot transfer results across subtyping and mutation prediction tasks against MIL baselines and recent pathology foundation models.
- 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.The paper tackles a genuinely important problem for multimodal pathology pretraining, and the deployment setup is practical since inference can still be unimodal. 2.The comparison study is reasonably solid, including both standard MIL baselines and several recent pathology foundation models. The gains are also fairly consistent across the few-shot settings. 3.The ablation studies are helpful and make it easier to see that both the architectural choices and the full loss design matter.
- 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 pathology report pipeline deserves a more careful leakage analysis. Even if explicit diagnostic labels are removed, reports can still contain subtype-related wording, biomarker mentions, or other highly label-proximal cues. 2.The empirical comparison is slightly incomplete. In particular, MIRROR [23] is one of the most related papers and should be included in the comparison. 3.The interpretability analysis is interesting, but I do not think the t-SNE plots by themselves establish that the model is really capturing synergistic information.
- 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?
I am overall positive about this paper because it addresses an important multimodal pathology problem and is supported by a fairly broad experimental study with strong results. On the other hand, some weaknesses do exist; for example, the report-based setup needs better leakage analysis, and one close multimodal comparison is missing.
- Reviewer confidence
Very confident (4)
- [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
Review #2
- Please describe the contribution of the paper
In this paper, the authors proposed a synergistic information disentanglement framework grounded in Partial Information Decomposition theory for slide representation learning. The proposed model includes a Synergistic Information Bottleneck (SIB) regulated via the designed objective, which minimizing redundancy while maximizing synergy, thereby distilling high-order cross-modality interactions.
- 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) The manuscript is clearly articulated and accessible. The figures and visualizations are both high-quality and illustrative. Furthermore, the study is underpinned by a thorough review of existing literature and a well-justified rationale for the proposed approach. (2) The authors perform extensive benchmark few-shot learning experiments on multiple cancer cohorts; the results validate its superiority for downstream analysis in comparison with an excellent choice of state-of-the-art supervised and self-supervised baselines. Detailed ablation studies are well-conducted for demonstrating the effect of each module and strategy design in the proposed method. (3) The design of maximizing the synergistic information of multimodal representation by mutual information is quite interesting, such a design is transferable to multiple scenarios in multimodal representation alignment and fusion. (4) The methodology is presented with clarity and appears reproducible. A significant strength of this work is the exclusive use of public datasets for experimental validation.
- 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.
I have identified no significant weaknesses in this work. However, I have provided several comments regarding the literature review, methodology details, and experimental settings in (12) to help improve the paper’s quality.
- 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?
No
- 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
(1) The literature review is comprehensive, covering most state-of-the-art methods like TANGLE, MADELEINE, and MIRROR. However, the recent work on [1,2,3] might provide additional context. The authors are encouraged to briefly discuss how Φ-Omni differs from these specific theoretical perspectives in the introduction. (2) The composite loss function defined in Section 2.4 introduces a hyperparameter γ to balance the contrastive alignment and the synergistic disentanglement. While the chosen value of γ=0.2 yields strong results, the manuscript would benefit from a brief discussion or citation regarding the selection strategy for this weight. For instance, was it tuned based on a specific validation set, or does it follow the common practice from similar works? A short explanation would enhance the reproducibility of the method. (3) The Section 2.1 text encoding pipeline mentions an ‘OCR-based pipeline with text cleaning.’ For the sake of completeness and reproducibility, it would be helpful if the authors could briefly describe the specific rules used for the ‘text cleaning’ step.
[1] Ming Y Lu et al. A visual-language foundation model for computational pathology [J]. Nature Medicine, 2024.[2] Tristan Lazard et al. Giga-SSL: Self-supervised learning for gigapixel images [C]. CVPR, 2023.[3] Yilan Zhang et al. Prototypical information bottlenecking and disentangling for multimodal cancer survival prediction, ICLR, 2024.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript demonstrates exceptional quality and clarity. Each component of the proposed framework is well-justified and has been empirically validated to enhance multimodal data alignment through self-supervised learning, as evidenced by extensive experiments on public datasets.
- Reviewer confidence
Very confident (4)
- [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
Review #3
- Please describe the contribution of the paper
This paper proposes Φ-Omni, a self-supervised learning framework for omni-modal WSI representation learning in computational pathology. The central contribution is a paradigm shift in multimodal pretraining, moving away from redundancy-driven contrastive alignment toward synergistic information disentanglement. Grounded in PID theory, the authors decompose the mutual information between the fused representation and the three modalities into three atoms: Redundancy, Uniqueness, and Synergy. Building on this decomposition, the author introduced the SIB module and GCP with ΦID objective. The authors validate Φ-Omni across eight diagnostic tasks on five public independent cohorts, demonstrating superior few-shot generalizabillity over SOTA supervised MIL and SSL baselines.
- 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.Clear motivation: The paper rightly points out that standard contrastive learning forces different modalities together, which keeps redundant info but loses complementary signals. Using PID theory makes perfect clinical sense, since diagnosis relies on how modalities combine, not just what they share. 2.Most slide-level SSL papers only do two modalities. This paper jointly integrates WSI, RNA-seq, and pathology reports. 3.The experimental setup strictly follows the TANGLE (k-shot linear probing, 10 random seeds, macro-AUC). The evaluation spans 8 tasks across 5 external cohorts, and the performance is robust.
- 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.Mirror[1] also uses a similar idea to fix modality collapse. This paper needs to explain exactly why your PID approach is fundamentally better and add an experimental comparison with Mirror. [1] Wang, T., Fan, J., Zhang, D., Liu, D., Xia, Y., Huang, H., Cai, W.: Mirror: Multi-modal pathological self-supervised representation learning via modality alignment and retention. arXiv preprint arXiv:2503.00374 (2025) 2.Training on around 2,000 TCGA slides is small compared to recent foundation models. Also, testing TCGA-trained models on CPTAC often suffers from site-specific biases. 3.Lack of ablation, please add Image+Gene only and Image+Text only to isolate the contribution of text reports.
- 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 main strength of this work is bringing PID into multimodal representation learning for computational pathology. However, this work lack of comparasion from Mirror and the pretaining datastets are small and limited in TCGA. My current recommendation is Weak Accept.
- Reviewer confidence
Very confident (4)
- [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
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
he reviewers are overall positive about this submission, which proposes a synergistic information disentanglement framework for omni-modal slide representation learning using WSI, transcriptomics, and pathology reports during pretraining while requiring only WSI at inference. The work addresses an important problem in multimodal computational pathology, is clearly presented, and shows strong few-shot performance across multiple cancer cohorts with useful ablation studies and comparisons to several MIL and pathology foundation model baselines. The main remaining concerns are the need for a more careful leakage analysis of pathology reports, clearer positioning and ideally comparison against the closely related MIRROR method, additional modality-combination ablations such as Image+Gene and Image+Text, and more details on text cleaning and hyperparameter selection. Overall, despite these limitations, the methodological idea is interesting, the experimental evidence is reasonably strong, and the paper is likely to make a valuable contribution if the authors adequately address these issues in the final version.