List of Papers Browse by Subject Areas Author List
Abstract
Acute ischemic stroke (AIS) requires time-critical decision-making, where inaccurate interpretation of neuroimaging findings can lead to irreversible disability. Diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps from magnetic resonance imaging (MRI) are central to detecting acute infarction, yet generating factually reliable radiology reports directly from 3D MRI remains challenging due to the difficulty of learning robust cross-modal alignments between volumetric images and clinical text. We propose paired image-domain retrieval and text-domain augmentation (PIRTA), a retrieval-augmented generation framework that improves report factuality by avoiding explicit image–text alignment. PIRTA retrieves clinically similar 3D DWI/ADC volumes using a pretrained 3D vision encoder and leverages their paired clinician-authored reports to ground large language model (LLM)–based report generation. Experiments on multi-institutional in-house data, a held-out external privacy-preserving cohort, and the public ISLES benchmark demonstrate that PIRTA achieves strong image-domain retrieval performance and consistently improves ischemic-territory accuracy, a clinically grounded surrogate for report factuality, compared to direct image-to-text baselines. These results indicate that retrieval-grounded generation provides a scalable and reliable paradigm for producing factually consistent radiology reports from complex 3D brain MRI. Source code is available at https://github.com/jhlee0619/PIRTA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1415_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jhlee0619/PIRTA
Link to the Dataset(s)
ISLES benchmark: https://www.isles-challenge.org/
BibTex
@InProceedings{LeeJun_Improving_MICCAI2026,
author = { Lee, Junhyeok AND Oh, Yujin AND Lee, Dahyoun AND Joh, Hyon Keun AND Kim, Minchul AND Sohn, Chul-Ho AND Baik, Sung Hyun AND Jung, Cheolkyu AND Park, Jung Hyun AND Choi, Kyu Sung AND Kim, Byung-Hoon AND Ye, Jong Chul},
title = { { Improving Factuality of 3D Brain MRI Report Generation with Paired Image-domain Retrieval and Text-domain Augmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces PIRTA, a retrieval-augmented framework designed to improve the factual accuracy of 3D brain MRI report generation. Instead of relying on explicit image-text alignment, PIRTA retrieves clinically similar image-report pairs to guide report generation, thereby reducing the complexity of cross-modal optimization. The framework combines robust image-domain retrieval with text-domain augmentation to achieve stronger factual grounding. It demonstrates higher ischemic-territory accuracy across diverse internal and external datasets, including public benchmarks, and highlights the benefits of large-scale self-supervised pretraining on unlabeled MRI data.
- 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 work focuses on 3D MRI data, which more accurately represent the volumetric nature of neuroimaging compared to 2D slice-based methods. 2.The paper is conceptually novel in deliberately avoiding explicit image-text cross-alignment, a step that is both computationally expensive and data-intensive in existing multimodal frameworks. 3.The proposed strategy allows the model to leverage large quantities of unlabeled data via self-supervised pretraining, resulting in a strong, generalizable 3D image encoder. 4.Extensive evaluation across multiple cohorts, including external and public datasets, supports the generalizability of the approach and its potential real-world utility.
- 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.As acknowledged by the authors, PIRTA’s reliance on retrieval similarity is a major limitation. In realistic clinical contexts, rare or atypical infarct patterns may not have sufficiently similar cases in the retrieval database, leading to factual degradation and limiting clinical applicability in its current form. 2.The dataset description requires greater clarity. The appearance of “multi-class” in Figure 4 is not previously defined; it seems to refer to cases with multi-territory infarcts, but this should be explicitly stated. 3.The terms “Pretrain No,” “Small,” and “Large” used in Tables 1 and 2 are inadequately explained. The paper should quantify how many samples were used in each pretraining scale to allow reproducibility and fair interpretation. 4.The data-splitting strategy within the internal cohort is not clearly detailed. It must be explicitly stated whether train, validation, and test partitions are separated at the patient level to avoid potential overlap or retrieval leakage between training and evaluation sets. 5.The baseline comparison raises concerns of fairness. PIRTA uses full 3D volumes and retrieval-based augmentation, while the baselines operate on 2D slices without retrieval grounding. This makes the comparison only partially controlled; a 3D baseline comparison would strengthen the paper’s empirical validity.
- 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 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?
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?
The paper presents an original and timely idea with meaningful clinical relevance, but the current version lacks methodological clarity and fair evaluation. I would not oppose acceptance after a strong rebuttal clarifying the data‑split strategy, comparison fairness, and retrieval limitations.
- 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 adequately addressed the comments I raised. Specifically, they added an additional experimental result for a fair comparison, clarified the dataset details, and acknowledged the limitations to be addressed in future work. Based on these revisions, I support acceptance of the paper.
Review #2
- Please describe the contribution of the paper
The paper addresses the challenge for 3D brain DWI/ADC report generation. Its major contribution is a retrieval-augmented report generation framework that bypasses the difficulty of modality-alignment, and is able to improves factual consistency in generating radiology reports.
- 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 framework is simple but well motivated. The use of techniques such as MAE and LoRA is appropriate for the specific task. 2.The method is evaluated on multiple independent cohorts, which provides useful evidence of its robustness and potential generalizability across different data distributions.
- 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 feature similarity used for retrieval is not sufficiently interpretable. There is a risk that the retrieval process may be driven by disease-irrelevant factors, which could reduce the reliability of reports generation.
- 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 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?
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 shows a good understanding of the key challenges in the proposed task and presents a method with a reasonable degree of novelty. Although there are some limitations in the design, these issue do not significantly detract from the overall contribution of the paper.
- Reviewer confidence
Somewhat confident (2)
- [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 have carefully reviewed the authors’ rebuttal and the revised manuscript. The authors have provided clear and convincing responses to the reviewers’ concerns, particularly regarding the data and the proposed framework. They have also incorporated a substantial amount of evidence based on performance metrics to support their arguments. Overall, the explanations are reasonable, and the additional analyses further strengthen the manuscript. Therefore, I recommend acceptance of the paper.
Review #3
- Please describe the contribution of the paper
The main contribution of this work is a retrieval-based framework for stroke report generation from DWI and ADC images. Rather than mapping images and text into a shared latent space, the method trains an image encoder to retrieve similar cases in the image domain. The reports of these nearest neighbors are then used as contextual input to a fine-tuned LLM, enabling grounded and clinically consistent report generation. The approach is validated on a private multi-institutional dataset and a public benchmark.
- 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 provides a clear and well-motivated problem formulation, highlighting the clinical importance of MRI, particularly DWI and ADC sequences, in stroke diagnosis and reporting.
The proposed idea is well justified. Stroke reporting is time critical and typically follows structured patterns, making it a suitable candidate for automated report generation.
The retrieval based approach is intuitive and aligns with clinical reasoning, as radiologists often rely on comparing current cases with previously seen examples. The study is supported by a large and diverse dataset, including 1,831 internal cases (993 stroke and 838 controls), as well as two external cohorts (n = 580 and n = 206, including the ISLES benchmark), which strengthens the validity and generalizability of the results.
The experimental setup is well described, and the results are presented in a clear and appropriate manner.
The authors conduct extensive experiments, including analysis of different levels of pretraining for the image encoder, and provide comparisons against strong baselines such as GPT 4o and LLaVA Med, demonstrating a thorough and competitive evaluation.
They also acknowledge the shortcomings of current metrics such as ROUGE or BLEU in assessing clinical correctness.
- 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 definition of evaluation classes raises concerns. The categorization into normal, anterior, deep gray, posterior, and multiclass is not aligned with standard clinical frameworks such as vascular territories (for example MCA, PCA, ACA). Additionally, the manuscript does not provide a detailed explanation or justification for these classes, and it remains unclear whether hemispheric differences are taken into account. This limits the interpretability of the reported results.
The discussion section is relatively limited and does not sufficiently analyze the implications of the results. In particular, the clinical quality of the generated reports is not discussed in depth. While stroke reporting is often structured (for example using ASPECTS), the current evaluation does not capture important aspects of clinical reporting and appears to rely primarily on the predefined coarse classes.
A significant portion of the manuscript is dedicated to methodological and mathematical explanations, particularly Sections 2.1 (Problem Definition and Cross-modal Learning Complexity) and 2.2 (PIRTA: Paired Image-domain Retrieval and Text-domain Augmentation). While these sections are clearly written, they could be more concise to allow for a more thorough discussion of results and limitations, especially given that the primary focus of the work is report generation.
Figures 2 and 4 are relatively small and visually dense, which makes them difficult to interpret.
- 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
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?
While the overall idea is sound and the training setup, as well as the dataset, appear appropriate, the evaluation design raises significant concerns. In particular, the simplification into coarse and clinically non-standard classes (normal, anterior, deep gray, posterior, multiclass) is not well justified and does not reflect established clinical frameworks such as vascular territories.
This simplification effectively reduces the evaluation to a classification problem, rather than properly assessing the quality of generated clinical reports. As a result, important aspects such as correctness, completeness, and clinical usefulness are not adequately captured.
The authors appropriately acknowledge that current metrics do not fully capture factual correctness in clinical reporting. However, these limitations are not sufficiently discussed in relation to their own evaluation strategy, nor is the chosen classification scheme adequately justified.
Consequently, although the methodological approach is promising, the current evaluation does not convincingly demonstrate the claimed improvements in clinical reporting performance.
- 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 main concern was the limited discussion of the clinical implications and limitations of the work. If the authors adequately address these points through the revised Discussion as stated in R3.W2, enlarge Fig. 2 and Fig. 4 for readability, and clearly state the current limitations regarding hemispheric distinction, then the concerns would be sufficiently addressed and the manuscript could be accepted on these points. Releasing the source code as stated in R1.7 and R2.7 would further strengthen reproducibility.
Author Feedback
We thank reviewers and respond by theme.
A. Territory scheme (R3.W1, R1.W2). The four-class scheme is a coarse vascular-territory label set consistent with AHA/ASA stroke imaging guidelines [Powers, Stroke, 2019]. It is not a per-vessel ontology. Anterior maps to ACA/MCA cortical territories, deep gray to lenticulostriate (MCA-deep), posterior to PCA/vertebrobasilar, and normal to no infarct. This grouping preserves class support while retaining clinically actionable localization. Finer ASPECTS/per-vessel evaluation is future work (R3.W2).
B. Data integrity and pretraining quantification (R1.W3, R1.W4). Train/validation/test are split strictly at the patient level. No patient appears in multiple splits, and the retrieval database is built from training-set embeddings only, so validation/test images never serve as retrieval candidates (to be stated in Sec. 3). “Pretrain Large” = MAE on 38,532 UK Biobank subjects (Sec. 3.1), “Small” = MAE on the internal cohort only (~1,831), and “No” = random initialization. Counts will be added to the table caption.
C. Baseline fairness (R1.W5). GPT-4o and LLaVA-Med are 2D image-conditioned models with no public 3D variant; we fed them the lesion-centered axial slice (mid-brain if unavailable) and identical prompts (Sec. 4.2). To address medical-domain specificity, we additionally evaluated Med-Gemma (a medical-domain VLM) under both slice strategies. Multi-class territory accuracy on internal/BRMH/ISLES: PIRTA 89.9/71.4/67.0; Med-Gemma v1 lesion 17.5/20.2/11.2, middle 24.4/21.4/18.4; v1.5 lesion 31.3/36.4/22.3, middle 32.7/36.9/21.8; best general-purpose VLM 29.4/21.7/36.4.Across all 2D baselines, PIRTA leads by at least +57.2/+34.5/+30.6 points (best 2D = Med-Gemma v1.5 on internal/BRMH, GPT-4o on ISLES). Per-class gap persists (posterior 95.4 vs <=34.8 internal, 87.4 vs <=50.6 BRMH). A 3D-supervised medical baseline remains the next step.
D. Retrieval interpretability and rare-case reliance (R2.W1, R1.W1). The retrieval encoder is supervised for ischemic-territory classification. Its embedding space reaches 94.04 / 71.21 / 70.87 mAP@1 (Table 1) and 93.06 / 69.23 / 68.10% multi-class territory accuracy (Table 2) on internal / BRMH / ISLES, indicating disease-relevant feature encoding rather than generic image similarity. Cosine similarity is a confidence cue. In Sec. 4.2, high-similarity retrievals align with correct territory generation, while low-similarity rare patterns mark cases needing clinical review. We agree this is a real limitation. We will make it explicit and treat hybrid retrieval with multimodal evidence as the next step.
E. Discussion depth (R3.W2). We adopted ischemic-territory accuracy because n-gram metrics (BLEU, ROUGE) miss clinically important errors, whereas territory is a clinically actionable report attribute for stroke localization and reperfusion triage (Sec. 4.2). The Discussion will be expanded to (i) clarify clinical implications of retrieval-grounded factuality, (ii) frame territory accuracy as a focused factuality surrogate rather than a complete report-quality metric, and (iii) delineate ASPECTS-style scoring and prospective expert evaluation as planned axes.
F. Specifics. (R1.W2) “Multi-class” in Fig. 4 means multi-territory infarcts (>=2 territories), which the caption will define. (R3.W3) Sec. 2.1-2.2 will be tightened to free space for the expanded Discussion. (R3.W4) Fig. 2 and Fig. 4 will be enlarged within page limits. (R3) Hemispheric distinction is not in the current label set and is left for per-vessel work. Pretraining scale matters. Tables 1-2 show +13.3/+14.5/+32.0 mAP@1 and +11.1/+16.1/+28.1 multi-class points (Large vs No) on internal/BRMH/ISLES, supporting large-scale SSL for cross-institution generalization.
G. Reproducibility (R1.7, R2.7). If accepted, we will release source code (training, inference, baselines, evaluation).
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.
The paper received a mixed review. Reviewers agree that the paper presents an original and sound idea with meaningful clinical relevance but the paper also suffers from methodolical clarity and the evaluation design. I would recommend an invite for rebuttal such that the authors can clarify concerns raised by the reviewers.
- 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 authors provide clarification on all points raised by Reviewers. Still some are left to address, like the retrieval interpretability and the 3D baseline comparison suggested by Reviewer 1, but with the details added (e.g., pretraining sizes of approximately 40,000 individuals’ data, adding Med-Gemma v1 and v1.5 to address concerns on baseline fairness) the paper has been greatly improved.
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 appreciated the clinical relevance and the soundness of the proposed method. The rebuttal has sufficiently addressed the concerns regarding fair comparison, dataset details, and limitations. Two reviewers who gave “weak reject” have decided to accept the paper. Therefore, an acceptance is recommended for the manuscript. Please make sure to incorporate the required additional information the final version of the paper.
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.
Reviewers agree that the work is clinically relevant and methodologically interesting, and the authors’ responses were also convincing. I therefore recommend acceptance.