List of Papers Browse by Subject Areas Author List
Abstract
In biomedical vision-language modeling, datasets are typically mined from scientific literature where compound figures are paired with short and often partially informative captions. Prior work on sub-figure extraction has been limited in both dataset size and generalizability. In addition, no existing effort has incorporated rich medical context in image-text pairs. We revisit data curation as a foundational component of effective biomedical representation learning. We address this by introducing a scalable curation pipeline that integrates transformer-based subfigure detection, subcaption extraction, and contextual text enrichment to improve alignment quality. We introduce Open-PMC-18M, a large-scale high-fidelity biomedical dataset comprising 18 million image-text pairs, spanning radiology, microscopy, and visible light photography. We train vision-language models on our dataset and perform extensive evaluation on 6 retrieval and 19 zero-shot classification tasks across three major modalities. The models trained on our dataset establish new state-of-the-art results in medical representation learning. We release our dataset, models, and code for reproducibility and further study into biomedical vision-language modeling and representation learning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2639_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/vectorInstitute/pmc-data-extraction
Link to the Dataset(s)
Open-PMC-18M: https://huggingface.co/datasets/vector-institute/open-pmc-18m
BibTex
@InProceedings{BagNeg_OpenPMC18M_MICCAI2026,
author = { Baghbanzadeh, Negin AND Islam, Mohammed Saidul AND Ashkezari, Sajad AND Dolatabadi, Elham AND Afkanpour, Arash},
title = { { Open-PMC-18M: A High-Fidelity Large Scale Medical Dataset for Multimodal Representation Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of this paper lies in the introduction of a large-scale, high-quality biomedical vision-language dataset and its validation through extensive experiments. Specifically, the authors curate Open-PMC-18M, a dataset consisting of approximately 18 million high-fidelity image-text pairs, constructed via a carefully designed pipeline that includes subfigure extraction, subcaption parsing, and contextual text summarization to improve image-text alignment quality. In addition, the paper demonstrates the effectiveness of this dataset by pretraining CLIP-style vision-language models on it and evaluating them across a wide range of downstream tasks. Experimental results show that models trained on Open-PMC-18M consistently outperform existing general-domain and medical-domain CLIP baselines on multiple retrieval and zero-shot classification benchmarks, highlighting the importance of both dataset scale and alignment quality in biomedical representation learning.
- 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 paper presents a well-designed and comprehensive data curation pipeline that effectively addresses a key limitation in existing biomedical vision-language datasets, namely the poor alignment between images and text. By combining subfigure extraction, subcaption parsing, and contextual text summarization, the authors significantly improve the quality of supervision at scale . In particular, the proposed subfigure detection model, trained on a large synthetic dataset, is a practical and valuable component that enables accurate decomposition of compound figures, which are a major source of noise in prior work. The paper also provides strong empirical validation across multiple downstream tasks, showing consistent improvements over existing medical CLIP-style baselines, suggesting that better data curation leads to more effective representation learning. Overall, the work has high practical relevance and is likely to be valuable to the research community due to the release of a large-scale, high-quality dataset and associated resources.
- 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.
While the paper presents a well-engineered and practical pipeline, the level of novelty regarding the data processing pipeline is relatively limited, as most components build upon existing techniques and the main contribution lies in the large-scale dataset. In addition, the evaluation mainly focuses on zero-shot and retrieval-based tasks. Although these are standard benchmarks for vision-language models, they may not fully demonstrate the effectiveness of the learned representations for broader downstream medical applications. Incorporating additional evaluations, such as linear probing or fine-tuning on established medical benchmarks (e. g. , MedMNIST), would further strengthen the empirical validation. Finally, while the results indicate consistent improvements, it would be helpful to more clearly disentangle the relative impact of dataset scale versus the proposed data curation strategy.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
My overall recommendation is based on the paper’s strong practical contribution and its relevance to an important problem in biomedical vision-language modeling. The work addresses a well-known limitation of existing datasets—namely, the poor alignment between images and text—and proposes a comprehensive and effective curation pipeline to improve data quality at scale. The resulting Open-PMC-18M dataset is substantial in size and appears to enable consistent performance gains across multiple benchmarks, suggesting that the approach is effective and likely to be useful for the community. At the same time, the paper places more emphasis on data curation and engineering rather than methodological innovation, as most components build upon existing techniques. Additionally, while the evaluation is solid in zero-shot and retrieval settings, further validation on downstream tasks (e.g., fine-tuning or linear probing) would strengthen the claims regarding broader applicability. Overall, I find this to be a valuable and well-executed data-centric contribution with clear practical impact. Despite some limitations in novelty and evaluation breadth, the strengths outweigh the weaknesses, which leads to my weak accept recommendation.
- 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
Review #2
- Please describe the contribution of the paper
The main contribution of the paper is the introduction of a new large-scale, high-fidelity dataset designed to support multimodal representation learning.
The paper proposes a data generation strategy involving generating synthetic multimodal data samples with subfigures, with the aim of improving retrieval and related downstream task 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.
A strength of the paper is the dataset contribution itself. Large-scale multimodal datasets can be valuable to the community, especially when they are designed to support representation learning across diverse biomedical image-text settings.
The paper explores an interesting synthetic data generation strategy, which provides a scalable way to enrich training data beyond what is readily available in existing resources.
The work includes retrieval-based evaluation, which is a relevant way to assess the usefulness of the learned multimodal representations.
- 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.
One point that would benefit from further clarification is how well the learned representations generalize to more realistic or external real-world scenarios. In Fig. 1, it is somewhat difficult to interpret the comparison fairly if the competing models were not trained on the proposed dataset. In that case, stronger recall or retrieval performance on this dataset may partly reflect training-data alignment rather than a broader advantage of the proposed approach. Clarifying this comparison would help readers better understand what conclusions can be drawn.
Another point concerns dataset composition and comparability. It would be helpful to know whether the other datasets used for comparison contain microscopy and radiology images at a similar scale and proportion. If the domain balance differs substantially, that may also affect how the retrieval comparisons should be interpreted.
While the dataset contribution is interesting, the algorithmic contribution currently could be clearer. In particular, it would be helpful for the paper to more explicitly explain why the synthetic generation strategy, especially the use of samples containing subfigures, is effective. For example, it would strengthen the paper to discuss whether the gains arise from a form of regularization, from increased multimodal diversity, or from something conceptually different than standard augmentation-like effects.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 leaned toward a weak accept because the large-scale multimodal dataset is potentially valuable to the community, but my score is reduced by addressable questions about the reported comparisons and the extent of the paper contribution beyond the dataset construction strategy.
- Reviewer confidence
Somewhat confident (2)
- [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
The paper introduces a large-scale biomedical image-text dataset (18 million pairs) with improved alignment via subfigure extraction and contextual text enrichment (subcaptions + summarized figure context). It also proposes a scalable curation pipeline that demonstrates improved performance in multimodal representation learning across retrieval and classification tasks.
- 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 paper addresses an pratical bottleneck in medical vision- language models: poor image-text alignment in pubmed central-derived datasets.
-
The idea of combining subfigure extraction + subcaption + contextual summarization is well-motivated.
-
The dataset scale (18M pairs) is substantial across multiple modalities. Experiments are relatively comprehensive with consistent improvements over existing datasets.
-
- 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 dataset is derived from PMC no further analysis of domian imbalance or bias. Furthermore, there is no discussion of how this affects downstream learning.
-
There is no direct metric for image-text alignment quality.
-
It is unclear how much gain comes from: subfigure extraction or subcaption or context summary.
-
- 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?
Dataset is derived from PMC open-access data without further characteriazation.
- 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
The detector is trained on synthetically generated compound figures may not capture real-world layout complexity.
- 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 novelty lies on integration. Key aspects such as text quality, dataset noise, and alignment fidelity are not sufficiently validated.
- 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.
The majority of reviewers provide positive recommendations. Reviews highlight a valuable community resource: Open-PMC-18M, an 18M high-fidelity biomedical image–text dataset with improved alignment via subfigure extraction, subcaption parsing, and contextual summarization. Experiments show consistent gains for CLIP-style pretraining across retrieval and zero-shot classification benchmarks.
