List of Papers Browse by Subject Areas Author List
Abstract
Deep learning methods face dual challenges of limited clinical samples and subtle inter-class differences among Diffuse Cystic Lung Diseases (DCLDs) when performing multi-class diagnosis from Computed Tomography (CT) imaging. Although Multimodal Large Language Models (MLLMs) exhibit promising potential for rare disease diagnosis, their performance is often compromised by insufficient domain-specific knowledge and the absence of reliable multimodal evidence, leading to increased hallucination risks. To address these limitations, we propose DCLDs-RAG, a multimodal retrieval-augmented generation framework for DCLDs diagnosis. Specifically, DCLDs-RAG integrates: (1) a disease-specific multimodal corpus of DCLDs cases and a DCLDs knowledge graph corpus; (2) a visual evidence retrieval learns angular margin-aware representations with high diagnostic discriminability; and (3) a knowledge–clinical evidence retrieval extracts disease-specific entities and relational patterns from the DCLDs knowledge graph as conclusive evidence to support diagnostic reasoning. An MLLM synthesizing retrieved evidence with query for final diagnosis. DCLDs-RAG is validated on the multi-center cohorts encompassing four types of DCLDs, achieving superior accuracy and generating evidence-based descriptions closely aligned with expert insights. Code is available at: https://github.com/Li-Haoqing/DCLDs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0323_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Li-Haoqing/DCLDs
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiHao_DCLDsRAG_MICCAI2026,
author = { Li, Haoqing AND Shi, Jun AND Chen, Xianmeng AND Jia, Qiwei AND Wang, Rui AND Peng, Yating AND Gao, Li AND Chen, Qianyi AND An, Hong AND Ouyang, Ruoyun AND Hu, Xiaowen},
title = { { DCLDs-RAG: Evidence-Grounded Multimodal Retrieval-Augmented Generation for Diffuse Cystic Lung Diseases Diagnosis } },
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
This paper introduces DCLDs-RAG, an evidence-grounded multimodal retrieval-augmented generation framework designed for the diagnosis of DCLDs. Addressing the imited clinical samples and subtle inter-class differences among DCLDs—which often lead to hallucinations in general MLLMs. DCLDs-RAG has been validated in multicenter cohorts, achieving superior accuracy and generating evidence-based descriptions closely aligned with expert insights.
- 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 work excels at generating evidence-based feedback and can provide specific clinical interpretations based on retrieved imaging features and patient history. This transparency is crucial for gaining the trust of clinicians.
- 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 authors mention “We extract representative CT slices from three views. . . “ but crucial implementation details are missing. Please clarify what are the quantitative criteria or clinical guidelines for defining a “representative” slice?
2.Please verify whether there is an error in Eq (6). When defining the retrieval objective for the top-k image-text pairs, the authors employed a arg min operation to process the cosine similarity scores. From a mathematical perspective, a higher cosine similarity value implies a shorter distance within the feature space, thereby indicating a higher degree of similarity between samples. Theoretically, to retrieve the top-k samples with the highest relevance, the arg max function should be used.
3.Section 3 only provides basic training parameters for the ResNet visual retriever, omitting crucial hyperparameters for the core RAG module. Given that Prompt design has a decisive impact on the MLLM’s output and its ability to mitigate hallucinations, it is strongly recommended that the authors provide the complete Prompt Templates and explain the rationale behind the core weight parameter settings in the path scoring functions (Eq 3-5).
4.Some Typos and Formatting Issues: 1) On page 2, “a visual evidence retriever expand. . . “ should be “expands”. 2) “** Hospital” on page 6. 3) The reference list lacks consistent formatting.
- 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?
While this paper proposes an innovative solution to challenging clinical diagnostic tasks, it still has some key shortcomings, which can be found in the major weakness section.
- 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
Proposed DCLDs-RAG, a multimodal RAG framework that integrates a domain-specific knowledge graph and clinical corpus to address data scarcity and mitigate MLLM hallucinations in rare disease diagnosis.
Introduced an angular margin-aware visual retrieval mechanism to capture subtle inter-class differences in CT images, ensuring high diagnostic discriminability for diverse lung diseases.
Synthesized multi-source evidence, combining visual features with relational patterns from knowledge graphs, to achieve superior diagnostic accuracy and provide expert-level, evidence-based reasoning.
- 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.
Hallucination Mitigation: Reduces the risk of misinformation by grounding MLLM outputs in a specialized DCLDs knowledge graph and multimodal clinical corpus, ensuring factual accuracy for rare diseases.
Enhanced Discriminative Power: Captures subtle inter-class differences in CT scans through angular margin-aware representations, effectively distinguishing between visually similar lung diseases despite limited samples.
Evidence-Based Interpretability: Boosts clinical trust by synthesizing retrieved visual and textual evidence into transparent diagnostic reasoning, providing explanations that closely align with expert medical insights.
- 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.
Reproducibility and Transparency Issues: The study omits critical implementation details, such as specific prompt configurations, and lacks a public code repository. This absence of open resources raises significant concerns regarding the reproducibility and verifiability of the experimental results.
Insufficient Dataset Scale and Outdated Baselines: The evaluation is conducted on a very small dataset of fewer than 200 samples, and the compared baselines are outdated, which undermines the statistical reliability and credibility of the findings. To ensure a fair and rigorous assessment, the framework should have been evaluated against modern agents/baselines (e.g., ReAct) using consistent, state-of-the-art commercial models.
Limited Methodological Novelty: The technical contribution appears incremental, as Multimodal RAG and Knowledge Graph integration are already well-established methodologies. The research primarily focuses on applying these existing techniques to a specific medical domain (DCLDs) rather than introducing a fundamental algorithmic innovation.
- 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?
My recommendation is primarily based on the paper’s limited methodological novelty and insufficient empirical rigor. While applying Multimodal RAG and Knowledge Graphs to DCLDs is clinically relevant, the technical contribution remains incremental, as these are well-established paradigms across various domains. The evaluation is significantly weakened by a small dataset of fewer than 200 samples and the use of outdated baselines; a more rigorous assessment against modern agentic frameworks like ReAct, using consistent state-of-the-art models, is necessary to establish true efficacy and credibility. Furthermore, the absence of critical implementation details such as specific prompt configurations and a public code repository, raises serious concerns regarding transparency and reproducibility, making the work currently unsuitable for publication in its present form.
- 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.
The authors claim that the contributions of their research lie in the dataset and methodology. However, contrary to what is stated in the abstract, RAG techniques do not include any agentic RAG methodologies. Furthermore, image model-based baselines appear to require more extensive investigation. Above all, the authors have not transparently disclosed their code or actual data samples via an anonymous repository or similar means. In this type of research, it is essential to transparently disclose artefacts during the review stage. Therefore, from a technical perspective, I believe this paper should be rejected.
However, if other reviewers or the ACs judge that the benefits of accepting the paper outweigh the concerns, given the specific characteristics of the domain, it would be acceptable to accept this paper.
Review #3
- Please describe the contribution of the paper
The authors propose DCLDs-RAG, a retrieval-augmented generation framework for diagnosing diffuse cystic lung diseases from CT images. The method combines a disease-specific multimodal corpus, a DCLDs knowledge graph, a visual evidence retriever that learns discriminative representations, and a knowledge-clinical evidence retriever that extracts structured disease-specific entities and relations to support reasoning. The authors validate the framework by showing improved diagnostic performance and outputs compared to the baselines.
- 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.
Methodological Innovation: The paper proposes a disease-specific multimodal RAG framework that combines visual retrieval, structured knowledge graph reasoning, and clinical evidence retrieval in a unified pipeline. This is interesting because the method goes beyond standard metrics, improving diagnostic accuracy and interpretability, which are all together especially relevant in rare disease settings.
External Validation and Ablation: The paper includes both an internal and an external test cohort from different hospitals, and it also provides an ablation study showing the contribution of visual evidence and clinical knowledge evidence. This is a positive aspect of the evaluation.
Clinical Relevance: The paper addresses a clinically meaningful and difficult diagnostic problem. Diffuse cystic lung diseases are rare and challenging to distinguish because of overlapping radiological appearances, so a framework that aims to reduce hallucination and provide evidence-based diagnoses is relevant for medical image analysis and clinical decision support.
- 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.
Limited Statistical Validation: Although the paper reports accuracy, precision, recall, and F1-score on internal and external test sets, it does not provide standard deviations, confidence intervals, or statistical significance tests to support the claim of superiority over baselines.
Insufficient Clarity in the Ablation Section: Table 2 can be followed with the help of the accompanying paragraph, but the Table alone would be clearer if the Visual Evidence and K-C Evidence columns explicitly indicated whether each component is enabled or disabled in each row. Also, it is not explicitly stated which MLLM was used for the ablation, even though the main results vary depending on the selected MLLM.
Minor Formatting Issues: Needs to be improved to enhance the paper’s readability. For example, the introduction says “a visual evidence retriever expand the angular margin,” which is a grammatical error, and Figure 4 is captioned as “Fig. 4.Fig. 4.Performance Comparison of three RAG methods exposed grammatical errors, and even the plot appears to include four settings: w/o RAG, GraphRAG, MedGraphRAG, and DCLDs-RAG, the wording says “Performance Comparison of three RAG methods”.
- 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?
My recommendation reflects a balance between recognizing the relevance of the clinical problem and identifying a few limitations that keep the paper closer to acceptance, and the reasons are the following:
Methodological Innovation: The paper targets a rare and clinically important diagnostic problem where standard discriminative models are likely to struggle because of limited samples and subtle inter-class differences. The proposed framework is methodologically interesting because it combines visual retrieval, structured knowledge graph evidence, and clinical evidence. Moreover, the figures and tables also help the reader understand the workflow of the proposed methodology more clearly.
However, the paper could be enhanced with a statistical evaluation. The reported performance gains are promising, but the paper does not provide standard deviations, confidence intervals, or significance testing, which is important given the relatively small and imbalanced dataset.
Overall, I find the paper interesting and potentially valuable for the MIC community, especially because of the rarity of the target disease and the explicit effort to ground multimodal reasoning in retrieved evidence.
- 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
Author Feedback
To R1: W1.CT Slice Selection. “Representative slices” were defined strictly following the cited clinical guidelines [9,22]. Two board-certified thoracic radiologists selected slices that capture the maximal cross-section of hallmark disease features (lentiform or irregularly shaped cysts with basal and subpleural predominance for BHD, and diffuse, uniformly sized round cysts for LAM, etc.). To capture these fine-grained structural nuances, slice thickness was strictly controlled to ~1.25mm. This rigorous standard ensures the extracted visual evidence is highly discriminative. W3.Hyperparameters & Prompt. For Eq. 3, α=1.0.For Eq. 4, w1=2.5 (eigenvector centrality) prioritizes core pathological hubs, while w3=1.0 (betweenness) captures bridging nodes crucial for differential diagnosis. We down-weight w2=0.5 (degree) to suppress non-specific common findings, and γ=0.2 penalizes excessively long paths. For Eq. 5, λ_1,2,3={1, 0.5, 2} heavily weighs specific clinical histories. Prompt is provided at the end. W2, W4.Eq.6 is corrected to argmax. All typos are fixed.
To R2: W1.Reproducibility. Prompt is provided at the end. Per rules, links are prohibited during rebuttal. Upon acceptance, we commit to fully open-sourcing our code and the extremely precious multi-center rare DCLDs dataset for full reproducibility. W2.Cohort Scale & Baselines. Given the ultra-rare nature of DCLDs, to our knowledge, our 214-case multi-center cohort is the largest globally. Regarding baselines, we respectfully disagree that they are outdated. We benchmarked against the latest, highly relevant SOTA methods: GraphRAG (2024) and MedGraphRAG (ACL 2025). Conversely, ReAct (2023) is older with weak domain relevance. Per the policy, we cannot introduce additional baselines during the rebuttal. W3.Our contribution transcends RAG adaptation. To our knowledge, DCLDs-RAG represents the first AI-driven diagnostic framework for DCLDs. Given the severe inter-class feature overlap in DCLDs, we introduce two architectural innovations: (1) Angular Margin-aware Visual Retrieval enforcing strict decision boundaries between entangled subtypes, and (2) Multimodal Evidence Synthesis (Eqs. 3, 4, 5), which systematically leverages high-specificity clinical evidence and knowledge priors. By grounding MLLM reasoning in verified clinical evidence, our framework substantially enhances diagnostic credibility and mitigates hallucination risks inherent in general models. These domain-driven designs are foundational to achieving expert-level reliability in the challenging clinical scenario of rare diseases.
To R3: W1.Statistical Validation. On the external set, DCLDs-RAG (GPT-5.2) achieved 68.8% acc (SD:0.46, 95%CI:[57.9,77.9]), significantly outperforming the best baseline, GraphRAG (55.0%, SD:0.50, CI:[44.1,65.4]) & MedGraphRAG (27.5%, SD:0.45, CI:[18.9,38.1]). Bootstrap CIs for acc gains: [3.8,23.8] (vs GraphRAG) & [30.0,52.5] (vs MedGraphRAG). Exact McNemar & paired permutation tests confirm significance vs GraphRAG (p=0.019 and p=0.021, respectively) and highly significant superiority over MedGraphRAG (p≈2.10e-9 and p≈5.00e-5). W2 W3.Table 2 will explicitly indicate component status. We confirm that GPT-5.2 was used for all ablation experiments. Minor issues have been corrected.
Prompt: Please perform a diagnostic evaluation based on the provided clinical context and multimodal evidence: 1.[Query CT Scans] Initial n patient slices. 2.[Retrieved Visual Evidence] Next k retrieved cohort CTs & descriptions. [Directive] Strictly ground diagnostic reasoning on the fine-grained margin constraints of the DCLDs below. 3.[Knowledge-Clinical Evidence (Graph RAG)] [Visual Evidence (Corpus RAG)] CT imaging phenotypes: …, Clinical profile: … [Diagnostic Inquiry] Based on the provided evidence, what is the diagnosis? A. BHD B. LAM C. PLCH D. LIP [Output] Answer: [Option]. [Disease] | Explanation: [Concise evidence-grounded reasoning]
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 generally positive but not fully converged assessment. Reviewers appreciated the clinical motivation and the overall retrieval-grounded design, but there are concerns about methodological detail, reproducibility, and the statistical support for the reported gains. In rebuttal, the authors should focus on clarifying these issues.
- 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 paper received two weak accepts and one reject, and the rebuttal addressed several of the main concerns. The authors clarified the slice selection protocol, prompt and parameter settings, ablation setup, and the statistical significance of the reported gains. While Reviewer 2 remains reject after rebuttal. I agree that reproducibility is still limited, especially without released code or data at review time. However, the authors did provide additional methodological and statistical clarification, and the paper includes comparisons to GraphRAG and MedGraphRAG, which are relevant baselines for this setting.
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 rebuttal substantially strengthened the submission: the authors clarified slice-selection criteria, hyperparameters, and prompts, and added statistical validation.
The review raised legitimate concerns about reproducibility and a mismatch between the abstract’s “agentic RAG” framing and the actual method; the latter is an overclaim the authors must correct, and the former is mitigated only by a commitment to release code and data upon acceptance.
On balance, I recommend acceptance, conditional on (1) revising the abstract to accurately describe the method without the “agentic” overclaim, and (2) a firm commitment to open-sourcing the code and dataset in the camera-ready version.
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.
The authors have addressed and clarified several concerns, for reproducibility data and code must be shared with the final paper.
