List of Papers Browse by Subject Areas Author List
Abstract
Acute ischemic stroke requires rapid delineation of infarct
core on diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) to guide time-critical decisions. Clinical urgency yields preliminary reporting as an early workflow summary centered on lesion location and extent, which enables report-conditioned multimodal segmentation. Paradoxically, the same urgency also makes preliminary reports frequently missing or inconsistently archived in retrospective cohorts, creating a key bottleneck for multimodal stroke segmentation. To address this gap, we present (i) Lesion-Grounded Self-Retrieval-Augmented Generation (LeG-RAG), a training pipeline that retrieves lesion-concordant reports using atlas-normalized lesion masks and synthesizes missing reports with a multimodal large language model without additional training, and (ii) LLMSwin, a report-conditioned multimodal segmentation architecture for brain MRI that extracts text features by combining the report with learnable prompts and feeding them into a frozen Llama-3-8B-Instruct model, and fuses these representations with multi-scale features from a SwinUNETR 3D backbone via two-way transformers. Experiments on a multi-institutional 3D stroke MRI dataset show that LeG-RAG achieves superior retrieval performance compared with state-of-the-art image-embedding methods across text-deficiency regimes, while remaining diagnostically consistent under ASPECTS-based reliability checks. LLMSwin achieves state-of-the-art segmentation across text-availability settings, outperforming image-only and prior multimodal segmentation baselines. Code is available at https://github.com/HeeseongEom/LeG-RAG.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6416_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HeeseongEom/LeG-RAG
Link to the Dataset(s)
N/A
BibTex
@InProceedings{EumHee_TextDeficient_MICCAI2026,
author = { Eum, Heeseong AND Lee, Junhyeok AND Jang, Joon AND Jang, Han AND Kim, Songsoo AND Choi, Kyu Sung},
title = { { Text-Deficient Multimodal Stroke Segmentation with Lesion-Grounded Self-Retrieval-Augmented Generation } },
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
this paper presented a pipeline for handling text-deficient multimodal stroke segmentation. To do so, the authors proposed LEG-RAG to retrieve lesion-matched reports and synthesize missing ones, and then use these reports in a report-conditioned segmentation model, LLMSwin. The experiments are fairly complete as well, with evaluations on report quality, segmentation performance and ablation studies.
- 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 is a meaningful problem to study. the paper targeted a practical clinical problem rather than an artificial benchmark setting, where missing or inconsistently stored reports are common in retrospective datasets.
- the paper does not just add text as an auxiliary input, but builds a full framework for retrieving, generating, and using missing reports. The lesion-grounded retrieval idea is also well motivated, since it uses atlas-aligned lesion information to find more clinically relevant examples than generic embedding similarity.
- 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 technical novelty is not extremely strong at the component level. Most of the main ingredients, such as retrieval-augmented generation, frozen LLM-based text encoding, prompt tuning, and transformer-based multimodal fusion, are already established ideas. So the contribution feels more like a well-designed integration of existing methods than a fundamentally new methodological innovation. 2 In addition, since the retrieval process depends on atlas-space lesion masks, the broader applicability of the framework to other settings may need more discussion.
- 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?
this is a solid paper with a well-motivated problem setting and a complete system design. I think the work is interesting and useful, especially from an application perspective. However, the innovation is more moderate, since much of the framework is built by adapting and combining existing techniques. In that sense, the paper is stronger as a problem-driven multimodal system paper than as a fundamentally new method paper.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper addresses the common issue of missing reports in stroke imaging cohorts by proposing the LeG-RAG framework. This framework uses lesion mask-driven retrieval and large model synthesis to fill in textual gaps. Simultaneously, the LLMSwin architecture is designed, utilizing frozen Llama-3 data to extract textual features and fusing them with the SwinUNETR 3D backbone network for segmentation.
- 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 LeG-RAG framework attempts to inversely complete the clinical context using atlas-normalized masks. This approach, which utilizes large language models (LLM) to achieve a closed loop in clinical reasoning, has practical application potential. 2.LLMSwin employs a dual-path Transformer to effectively align 3D spatial features with LLM semantics, achieving a good balance between parameter efficiency and segmentation performance.
- 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 MLLM and LLM models used in this article are general-purpose models, and neither is easily considered state-of-the-art, as they generally perform poorly on medical data. Also the BiomedCLIP. Consider using LLM models fine-tuned for medical data, such as Medgemma and Fleming-R1, and also consider more advanced general-purpose LLM models such as ChatGPT/Gemini/DeepSeek (if available), or open-source models like Llama4/Qwen3.5.Similarly, consider MLLM models such as MediX-R1/MedMO/Fleming-VL. 2.In spatial key-value retrieval, a cosine similarity threshold of delta = 0 is set. Due to the sparsity of 3D space, delta = 0 almost loses its filtering significance, which may lead to excessive load on the subsequent clinical context retrieval system and the introduction of a large number of noisy cases. Supplementary sensitivity experiments on delta (e.g., delta in [0, 0.05, 0.1]) are required to analyze its impact on retrieval efficiency and final segmentation accuracy. 3.The final testing phase inputs images and corresponding reports to output the final segmentation results, similar to Sam3.Furthermore, there are existing methods to fine-tune Sam3 on medical data (https://doi.org/10.48550/arXiv.2511.19046), and a comparison of these methods is requested. 4.In Table 1, LLMSwin achieves a segmentation accuracy (DSC) that matches the upper limit of 80% of real text when the text availability is only 20%. Please check and report the specific components of the dataset used in Stage/Laterality/Territory/Middle Cerebral Artery/HemorrhagicTransformation and analyze whether overfitting exists in the synthesized text.
- 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 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?
This paper has clinical application potential and a reasonable structural design, but it lacks professionalism in model medicine, lacks comparison with advanced models, and suffers from insufficient interpretability of its abnormally high performance conclusions. Furthermore, the experimental rigor has not yet reached the acceptable threshold.
- 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 #3
- Please describe the contribution of the paper
The main contributions of this paper are:
1.LeG-RAG - a lesion-grounded retrieval-andsynthesis framework for text-deficient stroke training. 2.LLMSwin - a reportconditioned volumetric multimodal segmentation architecture. 3.Comprehensive evaluation showing improved synthesized-report quality and robust segmentation under text deficiency.
- 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.Strong Performance: LLMSwin achieves state-of-the-art results across various text-availability settings.
2.Ablation Studies: The paper includes clear ablations showing that the combination of DSC and HD95 in the retrieval process is superior to using either metric alone.
3.Generalizability: Testing on two held-out sets (Test A and Test B) from multiple institutions suggests the model is robust.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Inference Efficiency: While the performance is strong, the use of a Llama-3-8B model—even when frozen—introduces significant computational overhead. A brief discussion on inference latency compared to traditional U-Net architectures would be beneficial.
- 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 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.
(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?
1.Clinical Relevance and Impact: The paper addresses a highly practical bottleneck in clinical AI: the missing or inconsistent archiving of preliminary reports in retrospective stroke cohorts. By enabling report-conditioned segmentation even when text is missing, the work aligns well with real-world clinical workflows.
2.Methodological Innovation: The LeG-RAG pipeline introduces a clever, spatially grounded retrieval strategy. Instead of relying on generic image embeddings, it uses atlas-normalized lesion masks to retrieve contextually relevant reports. The inclusion of the Clinical Context Score (CCS), combining Dice similarity and Hausdorff distance, is a well-motivated technical contribution for medical imaging.
- 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
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.
I align with the overall consensus reached by the reviewers