List of Papers Browse by Subject Areas Author List
Abstract
Medical vision–language pretraining increasingly relies on medical reports as large-scale supervisory signals; however, raw reports often exhibit substantial stylistic heterogeneity, variable length, and a considerable amount of image-irrelevant content. Although text normalization is frequently adopted as a preprocessing step in prior work, its design principles and empirical impact on vision–language pretraining remain insufficiently and systematically examined. In this study, we present MedTri, a deployable normalization framework for medical vision–language pretraining that converts free-text reports into a unified [Anatomical Entity: Radiologic Description + Diagnosis Category] triplet. This structured, anatomy-grounded normalization preserves essential morphological and spatial information while removing stylistic noise and image-irrelevant content, providing consistent and image-grounded textual supervision at scale. Across multiple datasets spanning both X-ray and computed tomography (CT) modalities, we demonstrate that structured, anatomy-grounded text normalization is an important factor in medical vision–language pretraining quality, yielding consistent improvements over raw reports and existing normalization baselines. In addition, we illustrate how this normalization can easily support targeted text-level augmentation strategies, including knowledge enrichment and anatomy-grounded counterfactual supervision, which provide complementary gains in robustness and generalization without altering the core normalization process. Together, our results position structured text normalization as a critical and generalizable preprocessing component for medical vision–language learning, and establish MedTri as a practical implementation of this approach. The source code is available at https://github.com/Arturia-Pendragon-Iris/MedTri.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0601_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Arturia-Pendragon-Iris/MedTri
Link to the Dataset(s)
https://zenodo.org/records/20690370
BibTex
@InProceedings{ChuYue_MedTri_MICCAI2026,
author = { Chu, Yuetan AND Ma, Xinghua AND Jin, Xinran AND Luo, Gongning AND Gao, Xin},
title = { { MedTri: Structured Medical Report Normalization for Enhanced Vision–Language Pretraining } },
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 presents MedTri, a lightweight, locally deployable normalization framework that converts free-text radiology reports into structured triplets for medical vision–language pretraining (VLP). It distills structured normalization capability from a cloud-based LLM (ChatGPT-5.1) into a compact biomedical encoder–decoder model via supervised fine-tuning on ~100k paired samples. The paper additionally proposes two optional text-level augmentation modules built on top of the normalized triplet: MedTri-K (medical knowledge expansion) and MedTri-C (anatomy-grounded counterfactual generation). The authors systematically evaluate the impact of this normalization strategy on downstream classification tasks across both X-ray (MIMIC-CXR, NIH ChestX-ray14, RSNA-Pneumonia) and CT (CT-RATE).
- 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 real and important preprocessing problem in medical VLP: radiology reports are long, stylistically heterogeneous, and often contain image-irrelevant content. 2.The proposed system achieves substantially lower computational cost than alternatives. 3.The framework is developed and evaluated across multiple modalities (both X-ray and CT) using large-scale, standard datasets, demonstrating a commendable engineering effort.
- 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.Insufficient novelty. The triplet-based report decomposition is highly reminiscent of MedKLIP (ICCV 2023, Wu et al. ), which decomposes reports into {entity, position, exist} triplets. It also includes an “entity translation” module that queries a medical knowledge base to expand disease entities into detailed visual descriptions — this is functionally very similar to MedTri-K. Without explicit comparison and differentiation, the claimed novelty of the triplet schema and the knowledge expansion module is substantially weakened. The comparison between their method and the related work mentioned is inadequate, which may further influence their novelty positioning. The paper cites MedFiLIP as a related work doing text normalization, yet does not include it as an experimental baseline. MedFiLIP uses GPT-based information extraction to produce fine-grained entities with severity, location, and category — a structure closely resembling the MedTri triplet. CT-GLIP already constructs organ-level image–text pairs and further introduces an abnormality dictionary to strengthen grounded contrastive learning. Relative to these papers, MedTri looks more like a new structured normalization schema plus a lightweight deployment recipe, rather than a fundamentally new paradigm. 2.Limited downstream task diversity. All downstream evaluations are zero-shot multi-label classification. There is no evaluation on retrieval, segmentation, or report generation — tasks that would more convincingly demonstrate the generalizability of the normalization. The discussion section acknowledges this but the limitation significantly narrows the empirical claims. 3.The improvements are modest. The improvements of MedTri over raw reports are relatively small (e. g. , 0.792 vs. 0.781 ACC on MIMIC 100% with SwinT), and no confidence intervals or statistical tests are reported for downstream tasks. It is difficult to assess whether the observed gains are reliably above noise, particularly in the 100% data regime. 4.The paper does not report the train/val split ratio (only mentioning 500 for testing) for the local text normalization model, does not describe how the 20 physicians were selected or inter-rater reliability, and does not discuss potential failure modes of the normalization (e. g. , hallucinated anatomical entities, information loss for rare conditions).
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper provides useful empirical evidence that structured text normalization benefits medical VLP. However, the work suffers from several issues. First, the triplet-based decomposition and knowledge expansion are not novel given MedKLIP. Also the baseline method comparison mentioned in the paper is insufficient. Second, the downstream evaluation is limited to classification only, the improvements are modest, and no statistical significance analysis is provided. Third, the “platform” framing oversells what is essentially a fine-tuned seq2seq model with two augmentation modules.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After reading the rebuttal, I maintain my original recommendation. The authors clarified the reason why they didn’t include or compare with relevant works, but my concern about novelty (components including triplet decomposition, knowledge expansion, structured report normalization) remains insufficiently addressed. If MedTri is reframed as an application study, the evaluation seems narrow both on modalities and tasks. Therefore, I maintain my decision.
Review #2
- Please describe the contribution of the paper
This paper introduces MedTri, a standardized framework that transforms messy free-text medical reports into structured triplets following the format of “anatomical entity: imaging description + diagnostic category”. It optimizes the text supervision of VLP, and delivers remarkable performance improvements across multi-modal scenarios and downstream 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.
1.The triplet graph designed in this framework accurately captures the basic semantic units for radiological reasoning, which efficiently eliminates irrelevant noise in medical reports that has no connection with medical images.
2.The researchers verified the efficacy of this method on two mainstream imaging modalities (X-ray and CT) as well as multiple public datasets.
3.MedTri achieves promising results under limited labeled data scenarios (1% and 10% data proportion). This is of great practical significance, considering the high cost and scarcity of annotations in medical artificial intelligence research.
- 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.Experimental outcomes show unstable performance of knowledge enhancement (MedTri-K) and counterfactual enhancement (MedTri-C) under different data scales. For instance, MedTri-K brings only trivial performance gains when using the full dataset.
2.This study only conducts evaluations on the classification downstream task. Testing over a broader range of downstream tasks would make the work more comprehensive and solid.
- 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.
(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?
On the whole, this is a valuable piece of research. It proposes anatomy-anchored normalization for structured reports and modular text enhancement strategies, and verifies that these designs greatly enhance vision-language alignment performance across various imaging modalities and model structures. For the above reasons, I recommend acceptance of this paper.
- 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.
My concerns have been addressed, so I maintain my score.
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is the proposal of MedTri, a lightweight and deployable text normalization platform designed to enhance medical vision-language pretraining (VLP). To address the stylistic heterogeneity and irrelevant content in raw radiology reports, MedTri converts free-text narratives into unified, anatomy-grounded triplets formatted as [Anatomical Entity: Radiologic Description + Diagnosis Category]. Building upon this structured schema, the authors further introduce two modular text-level augmentations: Medical Knowledge Expansion (MedTri-K) to enrich descriptions with standard radiologic signatures , and Anatomy-Grounded Counterfactuals (MedTri-C) to generate hard negative samples through localized perturbations for fine-grained multimodal alignment.
- 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 proposed [Anatomical Entity: Radiologic Description + Diagnosis Category] triplet schema effectively filters out stylistic noise and clinically irrelevant content while preserving essential morphological and spatial details for pretraining. 2.The structured format enables clever text-level augmentations without altering the core pipeline. Specifically, MedTri-K enriches knowledge for low-data regimes , and MedTri-C generates anatomy-grounded counterfactuals to improve fine-grained multimodal alignment.
- 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 framework relies on a lightweight local model to parse and generate the structured triplets. The paper lacks a robust error analysis (e.g., assessing the rate of hallucinated anatomies or missed negations by BioBart) and fails to evaluate how these inevitable extraction errors propagate and impact the downstream vision-language alignment. 2.Although the study notes that 20 board-certified physicians conducted a double-blind assessment using a 5-point Likert scale, the evaluation lacks crucial statistical transparency. The authors fail to disclose the exact number of samples evaluated by the experts and omit any inter-rater reliability metrics (such as Fleiss’ Kappa or Cohen’s Kappa). Without these fundamental statistical validations, the high clinical scores reported for MedTri lack the necessary robustness and remain unconvincing from a clinical perspective.
- 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?
This paper addresses a practical bottleneck in medical vision-language pretraining by proposing MedTri, a deployable text normalization platform. Its strengths lie in its high deployability (using a lightweight local model), multi-modality evaluation (X-ray and CT), and the inclusion of physician expert evaluation. However, the algorithmic novelty is somewhat incremental. Extracting structured triplets, utilizing knowledge dictionaries, and generating counterfactuals for hard-negative contrastive learning are well-explored concepts in broader NLP/CV communities. The primary contribution is essentially an effective engineering synthesis and domain adaptation rather than a methodological breakthrough. I recommend a Weak Accept because, despite the limited theoretical innovation, the systematic execution, solid empirical results, and practical utility for the medical imaging community are substantial.
- 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
Author Feedback
We thank all reviewers and AC for the constructive comments. We appreciate that the reviewers recognized MedTri as “a valuable piece of research” that “greatly enhances vision-language alignment” (R2), and as an “application study” with “high deployability,” “solid empirical results,” and “practical utility for the community” (R3). We are encouraged that R2 and R3 gave weak or strong accept recommendations; R1 concerns appear related to different interpretations of our intended positioning. Below, we clarify these concerns and explain revisions.
R1&AC: Contribution We would like to clarify that our intended positioning is: MedTri is an application study rather than a new methodology or paradigm. Our contributions are: 1.We validate text normalization as an independent factor for improving medical VLP. 2.MedTri provides a lightweight, locally deployable implementation. 3.MedTri supports easy modular text-side augmentations without changing the pipeline. We will revise “platform” to “deployable component” to avoid overstating the scope.
R1&AC: Compared baseline MedKLIP, MedFILIP, and CT-GLIP are highly relevant; however, they are not direct controlled baselines, because they combine additional VLP-specific designs, e.g. MedFILIP’s semantic similarity matrix and CT-GLIP’ segmentation mask and diagnosis contrastive loss. A full comparison can introduce additional factors, confounding the isolated effect of normalization and precluding a fair comparison. For MedKLIP, its triplets are extracted by RadGraph, which focuses on discrete clinical entities and labels rather than preserving anatomical radiologic descriptions. This is why we included RadGraph as the compared baseline.
R1&AC: Performance improvement Our conclusion is based on comparisons across datasets, scales, encoders, and modalities rather than a single setting. Wilcoxon signed-rank tests showed significant improvements of MedTri over raw reports and RadGraph on X-ray (p=0.000231 and 0.000194), and CT (p=0.03125), indicating that the gains are consistent, instead of random variation. These will be added in the revision, while confidence intervals will be added where applicable.
R2&AC: MedTri-K and MedTri-C The scale-dependent behavior is expected. MedTri-K is specially designed for low-resource settings, so its full-data gains can be saturated. MedTri-C uses anatomy-ground negatives and becomes more effective when enough samples support fine-grained discrimination. Statistical tests further show that their improvements are significant vs. MedTri, with p=0.00733 and 0.000245, respectively.
R1&R3&AC: Evaluation The reports were randomly split into train/val sets at 9:1 ratio. 20 board-certified physicians independently rated 50 randomly selected cases across. ICC was 0.983 for anatomical correctness and 0.990 for image groundedness, while Fleiss’ kappa was 0.656 and 0.634, indicating excellent reliability of averaged scores and agreement among physicians. Cohen’s kappa was not used because it is designed for two raters. These will be added in the revision.
R1&R3&AC: Extraction-error In our current study, this is addressed at two levels. 1) Physician evaluation directly assessed anatomical correctness and image groundedness, the two criteria most sensitive to hallucinated entities and information loss. The near-ceiling scores and strong reliability suggest that such errors were not systematic. 2) Models were trained directly on normalized reports, and the consistent gains indicate that residual errors, if present, did not dominate the VLP signal.
R1&R2&AC: Downstream task Our current choice of zero-shot classification was intended as a controlled probe of pretrained representation quality while keeping downstream heads and task-specific designs fixed. We therefore evaluate across modalities, datasets, encoders, and data scales. We will state this scope more clearly and discuss broader downstream evaluation as future work.
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.
MedTri proposes a lightweight text normalization framework that converts free-text radiology reports into structured triplets to improve medical vision-language pretraining (VLP), with optional knowledge expansion and counterfactual augmentation modules. Reviewers acknowledge the practical importance of tackling report heterogeneity, the low computational cost, multi-modality evaluation (X-ray and CT), and the strong performance in low-data regimes. However, the weaknesses are noted. Novelty is questioned, considering the triplet schema closely resembles MedKLIP, and missing baselines (e.g., MedFiLIP, CT-GLIP) weaken the claimed contribution. Downstream evaluation is limited to zero-shot classification only, without diverse tasks. Performance gains over raw reports are modest (e.g., 0.792 vs. 0.781 ACC on MIMIC), and no statistical significance or confidence intervals are reported. Additionally, there is no assessment of how extraction errors propagate to downstream VLP.
Issues to be discussed during rebuttal:
- Limited downstream task diversity (only classification; no retrieval, segmentation, or generation)
- Insufficient novelty and missing baseline comparisons (MedKLIP, MedFiLIP, CT-GLIP)
- Lack of statistical significance testing and confidence intervals for reported improvements
- Physician evaluation lacks inter-rater reliability metrics (Fleiss’ kappa / Cohen’s kappa) and sample size details
- Unstable performance of MedTri-K and MedTri-C across different data scales
- 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 rebuttal successfully resolves most critical issues. The authors have convincingly repositioned MedTri as an impactful application study that validates text normalization as an independent factor for medical VLP, with lightweight deployability and modular augmentations. The addition of statistical significance tests strengthens confidence in the reported improvements, and the detailed inter‑rater reliability metrics address transparency concerns about physician evaluation. However, two limitations persist. First, downstream evaluation remains confined to zero‑shot classification; the absence of retrieval, segmentation, or generation tasks narrows the demonstration of generalization. The authors acknowledge this as future work, which is acceptable given the paper’s scope. Second, the modest absolute gains in full‑data settings, while statistically significant, raise questions about practical clinical impact, though the low‑data regime improvements are more compelling. Reviewer #1’s novelty concerns are partially mitigated by the clarified positioning, but researchers seeking fundamental methodological advances may still find the contribution incremental. Overall, the practical utility, solid empirical execution across X‑ray and CT, and responsive rebuttal justify acceptance.
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.
There is 1 Reject and 1 Accept after rebuttal. R3 did not update the final recommendation and was initially positive. Reviewers generally agree that the work addresses a practical problem in medical VLP and demonstrates utility across datasets and modalities. The rebuttal clarified the intended positioning of the work, provided additional details on statistical analysis and physician evaluation, and addressed several concerns regarding the augmentation modules and normalization quality. While concerns about novelty relative to prior structured-report approaches remain, I view the contribution primarily as a useful and well-executed application study. I recommend acceptance.