List of Papers Browse by Subject Areas Author List
Abstract
Pathology foundation models learn morphological representations through self-supervised pretraining on large-scale whole-slide images, yet they do not explicitly capture the molecular state of the tissue. Spatial transcriptomics technologies bridge this gap by measuring gene expression in situ, offering a natural cross-modal supervisory signal. We propose MINT (Molecularly Informed Training), a fine-tuning framework that uses spatial transcriptomics as molecular supervision for pretrained pathology Vision Transformers. MINT appends a learnable ST token to the ViT input to encode transcriptomic information separately from the morphological CLS token, preventing catastrophic forgetting through DINO self-distillation and explicit feature anchoring to the frozen pretrained encoder. Gene expression regression at both spot-level (Visium) and patch-level (Xenium) resolutions provides complementary supervision across spatial scales. Trained on 577 publicly available HEST samples, MINT achieves the best overall performance on both HEST-Bench for gene expression prediction (mean Pearson r = 0.440) and EVA for general pathology tasks (0.803), demonstrating that spatial transcriptomics supervision complements morphology-centric self-supervised pretraining.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5167_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
HEST dataset: https://github.com/mahmoodlab/hest
EVA evaluation framework: https://github.com/kaiko-ai/eva
BibTex
@InProceedings{LeeMin_MINT_MICCAI2026,
author = { Lee, Minsoo AND Kim, Jonghyun AND Yun, Juseung AND Yu, Sunwoo AND Jang, Jongseong},
title = { { MINT: Molecularly Informed Training with Spatial Transcriptomics Supervision for Pathology Foundation Models } },
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
- The paper proposed MINT (Molecularly Informed Training with Spatial Transcriptomics Supervision)
- the ST and CLS tokens capture complementary information— the ST token specializes for molecular signals while the CLS token retains morphological transferability—and that combining both yields consistent improvements in a backbone-agnostic manner.
- Under the official HEST-Bench and EVA evaluation framework, MINT achieves the best overall performance on both benchmarks (mean Pearson r=0.440 and EVA average 0.803).
- 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.
- The establishment of a method for fine-tuning on spatial transcriptomics datasets is highly commendable.
- The validity of introducing ST tokens and the overall framework is demonstrated through thorough evaluation.
- The motivation is clear, and efforts toward fine-tuning foundation models are highly important.
- 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.
To what extent does learning ST signals contribute to performance? I am interested in the performance when using only $L_{DINO}$, as well as when combining $L_{DINO} + L_{distill}.
- How effective is the positive-only supervision that ignores zeros? For observed genes, zero values can still represent an important signal of low expression, so discarding them may result in a loss of useful information.
- How does the performance compare to models trained from scratch for gene expression prediction?
- 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 does not provide sufficient information for reproducibility.
- 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?
As noted in the major strengths, I agree with the concept of the study and believe it is a valuable contribution to the community.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
As noted in the major strengths, I agree with the concept of the study and believe it is a valuable contribution to the community.
- 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.
I found the rebuttal convincing. Methods for tuning foundation models with ST are still underexplored. I believe this work will serve as a useful baseline for the community.
Review #2
- Please describe the contribution of the paper
1) This paper proposes a fine-tuning framework for spatial-transcriptomics (ST) prediction using only HE patches. It augments the traditional DINO framework with two additional supervision losses. 2) The paper further shows the morphological transferability of the pretrained backbone through the authors’ reported experiments through both HEST-Bench and EVA.
- 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 studies a meaningful problem with clear practical relevance. 2) The proposed framework is technically coherent and clearly presented. Its method is easy to follow, and the overall design is internally consistent. The paper is generally well organized, and the training objectives are described in a relatively clear way. 3) The design of the ablation study well supports the authors’ claim that the ST token and distillation strategy play complementary roles. 4) The distillation-based design for mitigating catastrophic forgetting is interesting and may provide useful guidance to the community for incorporating new supervisory signals into pretrained pathology foundation models.
- 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 most serious issue is the lack of direct comparison with prior spatial transcriptomics-based models. Although the paper cites several prior works ([11-14]) on histology-based gene expression prediction and related ST-supervised learning, these methods are not included as direct experimental baselines. 2) The HEST-Bench improvements are not entirely surprising because in-domain data with labels were used during training. This makes the benchmark more favorable to the proposed model than to other foundation models. An additional ST-prediction benchmark would be needed for a fairer comparison. 3) The core idea, namely adding an extra supervision loss and using distillation to stabilize fine-tuning, is somewhat trivial and seems more like a careful combination of existing techniques in the pathology-ST setting. 4) The paper’s central claim of avoiding catastrophic forgetting is interesting, but it is supported only indirectly. While the reported results are suggestive, the paper does not provide sufficiently strong mechanistic evidence that the pretrained morphological feature space is truly preserved during ST-guided fine-tuning. A feature-space visualization would be helpful.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
See weakness
- Reviewer confidence
Very confident (4)
- [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 authors’ feedback addressed my concerns regarding the selection of comparison methods. I also find it convincing that this work can serve as a fine-tuning framework for pathology foundation models.
Review #3
- Please describe the contribution of the paper
This paper presents a systematic approach to fine-tuning pretrained pathology ViTs using Spatial Transcriptomics supervision. It introduces a learnable ST Token into the Transformer input to encode molecular data, decoupling it from morphological representations to prevent catastrophic forgetting.
- 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 task definition and cross-modal signal selection are pioneering. Reverse-supervising morphological H&E models with molecular-level gene expression embodies true interdisciplinary innovation. 2) The authors successfully fine-tuned a massive 1.1B parameter foundation model by integrating dual-resolution transcriptomics data (spot-level Visium and patch-level Xenium). The quantitative results on both HEST-Bench and EVA are exceptionally thorough and convincing.
- 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) Lack of Qualitative Visualizations: For a representation learning paper claiming to successfully decouple morphological (CLS) and molecular (ST) features, the total absence of qualitative visualizations in the experimental section is a critical flaw. Without UMAP/t-SNE projections of the latent space or attention heatmaps demonstrating what histological structures the ST token focuses on, the model remains a black box, making it difficult to intuitively verify the decoupling claims. 2) The core algorithmic modification (adding a learnable token to prevent forgetting) is relatively simple from a general deep learning perspective, though highly effective in this specific application.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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
Regarding the lack of qualitative visualizations: In the rebuttal, please explain how you guarantee that the ST and CLS tokens are genuinely decoupled in the latent space without relying on visual proofs. Furthermore, please thoroughly revise the bibliography to conform strictly to the required Springer LNCS / MICCAI format.
- 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?
This paper presents a brilliant interdisciplinary framework that decouples molecular and morphological learning to supervise a massive 1.1B parameter foundation model. The workload is substantial, and the results are compelling. However, as a representation learning paper, the complete lack of qualitative visual proofs (UMAP/attention maps) to substantiate the feature decoupling claims is a significant weakness in the submitted manuscript. The reference formatting issues are a minor flaw that should be easily rectified. Satisfactory explanations regarding the feature decoupling mechanisms and a commitment to fix the formatting could elevate this 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 authors provided new quantitative metrics (CKA and cosine similarity) in the rebuttal to prove feature decoupling. While the numbers look promising, MICCAI guidelines explicitly forbid relying on new independent experimental results for the final decision. Furthermore, without qualitative visualizations (UMAP/heatmaps), the interpretability claims in the original manuscript remain unverified. I maintain my Weak Accept score based strictly on the submitted manuscript.
Author Feedback
Comparison with prior ST prediction methods (R2) MINT is designed as a foundation-model fine-tuning framework, not a task-specific ST predictor; ST prediction is used as supervision for representation learning. We omitted ST-head-based evaluation from the main paper because our main evaluation follows the official HEST-Bench frozen-feature PCA+Ridge protocol, which assesses MINT as a foundation-model representation rather than a task-specific ST predictor. Nevertheless, we agree that direct ST prediction performance is informative. Under the matched HEST-Bench protocol reported in STFlow, MINT achieves 0.4400 with frozen PCA+Ridge features, competitive with specialized ST predictors without using its ST head. In the ST-head evaluation, after 50 epochs of head fine-tuning, MINT reaches 0.4737, exceeding the reported STFlow result of 0.448 and other baselines: STNet 0.310, HisToGene 0.260, BLEEP 0.400, and TRIPLEX 0.432. Evaluation fairness (R2) HEST-1k is a heterogeneous meta-collection of independent public ST cohorts, not a single homogeneous in-domain dataset, and HEST-Bench uses held-out benchmark samples. Thus, the gain cannot be explained simply by sample overlap or a single in-domain training distribution. We agree that ST supervision is naturally aligned with ST prediction; therefore, we additionally evaluate EVA, which spans nine distinct non-ST datasets/tasks. MINT improves HEST-Bench while achieving the best EVA average, indicating that the learned representation generalizes beyond ST prediction.
Quantitative evidence for CLS token preservation and ST token specialization (R2, R3) We added an independent feature-similarity analysis on 10,000 TCGA patches balanced across pan-cancer types. MINT-CLS remains highly aligned with frozen H-optimus-0 CLS features, with CKA 0.9755 and mean cosine similarity 0.9330, supporting preservation of the pretrained morphology-centric representation. By contrast, MINT-ST is much less similar to frozen-CLS features, with CKA 0.6632 and mean cosine 0.2615, and is also distinct from MINT-CLS, with CKA 0.6787 and mean cosine 0.2829.These results provide quantitative evidence that MINT preserves morphology-oriented CLS features while learning a specialized ST representation.
Ablation of DINO, distillation, and ST supervision (R1, R2) We added two objective ablations. DINO-only yields 0.3748 on HEST-Bench and 0.761 on EVA, while DINO+distillation without ST losses improves to 0.4117 and 0.783, still below frozen H-optimus-0 baseline 0.4150 and 0.793.Full MINT reaches 0.4400 and 0.803.Thus, distillation helps preserve pretrained representations, while ST supervision is key to improving molecular informativeness and full MINT maintains strong general pathology performance.
Clarification of positive-only supervision (R1) We clarify that positive-only supervision does not discard all zero-valued gene entries. It is applied at the 14×14 patch level: we skip the ST loss only when all genes in a patch are zero. Thus, zero values within informative patches are still retained. Since the term “positive-only” may be misleading, we will refer to it more precisely as “all-zero patch filtering” and clarify this detail in the revision.
Novelty (R2, R3) We agree that learnable tokens and distillation are established techniques. The main novelty of MINT is not these components themselves, but using ST data as supervision for pathology foundation model adaptation, rather than treating ST prediction as a standalone task. MINT shows that ST data can serve as an additional supervision axis for pathology foundation models, molecularly informing morphology-centric representations while preserving transferability. Our results show gains in both molecular informativeness on HEST-Bench and general pathology performance on EVA.
Reproducibility and formatting (R1, R3) We will release the code/checkpoints upon acceptance and correct the bibliography formatting to follow the MICCAI/LNCS style.
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.
While the task addressed is important and timely and a consistent ST-supervised fine-tuning method is presented to deliver high performance, some concerns still remain and the concerns should be addressed in the following areas: 1) lack of a more comprehensive direct comparison to existing literature on existing histology-to-ST methods or ST-supervised fine-tuning methods for digital histology, 2) potentially unfair evaluation as the benchmark uses histology-to-ST matching trained using in-domain ST labels, 3) no sufficient evidence of how the introduced ST/CLS token design is effective to preserve morphological representation of ST images and reduce the catastrophic forgetting issue, and 4) insufficient ablation studies to demonstrate the effects of ST supervision, the usage of DINO loss, distillation, positive-only regression and training from scratch. I encourage the authors to add the missing evidence to strengthen the results in 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.
clear acceptance for this paper
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.
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 #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.
All reviewers recognized the importance and timeliness of using spatial transcriptomics supervision to improve pathology foundation models, with strong reported results on both HEST-Bench and EVA. The rebuttal satisfactorily clarified the framing of MINT as a foundation-model fine-tuning method rather than only an ST prediction model, addressed concerns about baseline comparisons, fairness, token preservation/specialization, and key ablations; remaining limitations on qualitative visualization and bibliography formatting are not sufficient to outweigh the contribution.