Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Diffuse Cystic Lung Diseases (DCLDs) comprise a group of rare pulmonary disorders with heterogeneous morphological patterns and overlapping imaging manifestations, posing a significant challenge for differential diagnosis. Although Multimodal Large Language Models (MLLMs) have shown promising zero-shot capabilities for DCLDs recognition, their limited context length prevents them from processing the entire CT scan, and the resulting incomplete spatial perception often leads to prediction hallucinations, characterized by poor accuracy and interpretability. To address these problems, we propose an interpretable MLLM-driven diagnostic reasoning framework for DCLDs inspired by the clinical workflow. To enable effective utilization of CT images, we design a Global Spatial Perception (GSP) strategy that projects critical 3D lesion characteristics into compact 2D representations, preserving essential spatial distribution and morphology information while substantially reducing input context. Furthermore, we introduce a Multistep Self-Verification Reasoning (MSVR) mechanism that guides the model to identify and validate discriminative visual evidence from representative axial slices, iteratively refining diagnostic probabilities and enhancing interpretability. We construct a large-scale, multi-center dataset comprising diverse DCLDs subtypes and conduct extensive experimental evaluations. Our method achieves a diagnostic accuracy of 82.5%, significantly outperforming state-of-the-art approaches, while ablation studies confirm the contribution of each key component. These results demonstrate the feasibility and clinical potential of MLLMs for reliable DCLDs diagnosis. Code available at https://github.com/JQW2006/dcld-gsp-msvr.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1577_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/JQW2006/dcld-gsp-msvr

Link to the Dataset(s)

N/A

BibTex

@InProceedings{JiaQiw_Multimodal_MICCAI2026,
        author = { Jia, Qiwei AND Li, Haoqing AND Shi, Jun AND Chen, Xianmeng AND Wang, Rui AND Peng, Yating AND Gao, Li AND Chen, Qianyi AND An, Hong AND Ouyang, Ruoyun AND Hu, Xiaowen},
        title = { { Multimodal Large Language Model-Driven Self-Verification Reasoning for Interpretable Diagnosis of Diffuse Cystic Lung Diseases } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper’s main contribution is its novel hybrid formulation, iterative reasoning paradigm, strong interpretability, and clinically aligned design, supported by convincing experimental validation. It addresses key MLLM limitations (context length, hallucinations) in 3D CT via clinically-inspired GSP (coronal rendering with slice-index map) and MSVR (hypothesis-driven retrieval + self-verification), enabling interpretable reasoning akin to radiologists. The results outperform the existing works by acheiving 82.5% accuracy, F1= 80.0%, with ablations showing GSP (+32.7% acc) and MSVR (+10.6%) impact.

  • 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.Novel integration of global and local reasoning for volumetric CT analysis The paper introduces a hybrid diagnostic formulation that combines global summarisation of 3D CT data with localised slice-level verification. By first constructing a compact global representation and then refining decisions using selected slices, the method achieves a balance between holistic understanding and fine-grained evidence analysis, resulting in both efficiency and clinical meaningfulness. 2.The proposed model works in steps instead of making a single decision by introducing a feedback-driven inference mechanism. The novelty lies in treating diagnosis as a dynamic reasoning process rather than a static prediction task, which is particularly interesting because it closely mirrors how clinicians refine decisions based on evidence. 3.The framework provides explicit intermediate outputs, including which slices it used as evidence, how its decision changed over time and a confidence score. This is a strong contribution because interpretability is often a bottleneck in deploying AI systems in healthcare. Unlike black-box models, this approach enables traceability and interpretability, improving clinical trust and facilitating human-AI collaboration. 4.The proposed global summarisation mechanism allows the model to process hundreds of CT slices without exceeding computational limits. This is an effective and practical solution to a common challenge in volumetric imaging. 5.The model performs better than existing deep learning methods. The paper also includes ablation studies by demonstrating performance degradation when removing either the global context or the reasoning loop, supporting the design choices.

  • 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 study is conducted on a relatively small cohort (~217 cases), with an uneven distribution across disease categories. This introduces two key limitations: the risk of overfitting and bias toward dominant classes, and results may not generalise well to new data.
    2. The system uses existing large models (like Gemini-3-Pro) without training them specifically for this task. As a result, the reasoning process is guided by prompts rather than learned policies, and slice selection and verification may be suboptimal or inconsistent.
    3. The paper focuses primarily on aggregate performance metrics, with insufficient investigation into failure cases and potential systematic biases.
    4. The paper does not provide inference time comparisons, resource requirements, or scalability analysis to assess practical feasibility in clinical workflows.
    5. The paper does not sufficiently analyse the trade-off between compression and diagnostic fidelity, which is critical in medical imaging tasks.
    6. While the method is compared against standard deep learning approaches and relatively old 3D CNN models. It requires comparing with state-of-the-art 3D CNN/Transformer models, or recent hybrid global-local architectures.
  • 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 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?

    The paper presents a promising and well-motivated approach that combines global CT summarisation with iterative, evidence-based reasoning, offering improved interpretability and a workflow aligned with clinical practice. The idea of closed-loop verification is interesting and supported by clear ablation studies, suggesting that both global context and iterative reasoning contribute to performance. However, the overall impact is somewhat limited by dataset size and imbalance, reliance on non–task-specific models, and the absence of efficiency and failure-case analysis. These factors raise concerns about robustness, reproducibility, and real-world applicability.

  • 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

    This paper presents an interpretable MLLM-based framework for diagnosing diffuse cystic lung diseases (DCLDs) from chest CT. The method combines Global Spatial Perception (GSP) to compress 3D lesion distribution into a compact 2D representation, and Multistep Self-Verification Reasoning (MSVR) to iteratively retrieve representative slices, verify fine-grained morphological evidence, and update diagnostic probabilities. The paper also contributes a multi-center DCLD dataset with pixel-level cyst annotations and demonstrates strong gains over supervised and prior medical-MLLM 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.
    • The paper tackles a clinically important and challenging rare-disease diagnosis problem with clear practical relevance.
    • The proposed framework is novel and well motivated, as it mirrors the radiologist workflow of combining global distribution assessment with local feature verification. The paper offers a solution to the context-length limitation of MLLMs for volumetric CT through GSP, while MSVR improves interpretability and reasoning transparency.
    • The experiments are comprehensive, which support the value of the proposed design.
  • 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 strongest results rely on proprietary MLLM backbones, which may reduce reproducibility.
    • The dataset size remains modest, especially for a 4-class rare-disease setting, so broader validation would strengthen the claims.
    • The framework depends on accurate upstream segmentation and structured preprocessing, which may complicate deployment.
  • 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 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?

    The paper is novel in its integration of compact global 3D perception with iterative self-verification reasoning, and it shows meaningful empirical improvements over strong baselines. The work is also clinically well motivated and provides interpretable outputs, which is especially valuable in medical AI.

  • 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 #3

  • Please describe the contribution of the paper

    This paper addresses the differential diagnosis of Diffuse Cystic Lung Diseases (DCLDs). The authors propose an interpretable MLLM-driven diagnostic reasoning framework comprising two core components:  Global Spatial Perception (GSP): Projects 3D cyst spatial information from volumetric CT into a single compact 2D rendering with an aligned slice-index axis, circumventing the limited context window of MLLMs while preserving distribution patterns and morphological cues. Multistep Self-Verification Reasoning (MSVR): Iteratively performs hypothesis-driven slice retrieval, 2D morphological evidence extraction, and Bayesian-style posterior updates until a confidence threshold is met or maximum iterations are reached. Dataset contribution:A multi-center, pixel-level annotated DCLDs dataset (217 cases across 4 subtypes) with a dedicated cyst segmentation model.

  • 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.

    S1.Well-motivated clinical problem with a cognitively aligned framework design. The clinical challenge of DCLDs differential diagnosis is clearly articulated: rare disease with limited data, overlapping cystic phenotypes requiring both global spatial reasoning and local morphological analysis. The two-stage GSP→MSVR pipeline elegantly mirrors the radiologist’s cognitive pathway of “first assess overall spatial distribution, then locally verify key morphological features” [Gupta et al., ERJ 2015, Ref. 5 in paper]. This clinical grounding is a notable strength over generic MLLM prompting approaches. S2.Novel and effective GSP representation. The idea of compressing 3D cyst distribution into a single surface-rendered image with a slice-index navigation axis is creative and pragmatic. The ablation study strongly supports its value: removing GSP causes a dramatic accuracy drop of 32.7 percentage points (82.5% → 49.8%). Additionally, the finding that ConvNeXt-S trained on 3D renderings (76.4%) outperforms all standard 3D volumetric models (≤65.0%) independently validates the discriminative power of this representation, even outside the MLLM framework. S3.Self-verification mechanism with demonstrated hallucination correction. The MSVR closed-loop design—where the model actively retrieves evidence to confirm or refute its current hypothesis—draws on established reasoning paradigms (CoT [Ref. 20], Reflexion [Ref. 23], ReAct [Ref. 24]) but applies them in a medically meaningful way. The qualitative example in Fig. 4 is compelling: standard Gemini-3-Pro produces a hallucinated BHD diagnosis biased by clinical metadata priors, while the proposed framework retrieves axial slices revealing morphological evidence inconsistent with BHD and correctly revises the diagnosis to LAM. S4.Comprehensive experimental design. The evaluation includes: (i) multi-center validation (internal + external cohorts); (ii) systematic comparison against 6 supervised baselines and 2 medical MLLMs; (iii) parameter sensitivity analysis for slice budget M and max iterations T; (iv) a well-structured ablation study isolating the contribution of each component. The cross-center generalization gap is notably smaller for the proposed method than for supervised baselines, suggesting robustness.

  • 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.

    W1.Dataset Size Still Modest Although the dataset is collected from multiple centers, its overall size remains relatively limited, comprising 217 cases in total, with only 80 cases in the external cohort and just 23 PLCH cases among the four subtypes. Such small subgroup sizes may introduce substantial variance and limit the statistical reliability of the reported results. Moreover, the manuscript does not provide confidence intervals or statistical significance analyses, making it difficult to assess the robustness and stability of the performance improvements. W2.Lack of Direct Comparison with Radiologists The study does not include a direct comparison with radiologists or pulmonologists. Without benchmarking against human expert performance, the claim of “physician-level reasoning” is not sufficiently supported. W3.Hallucination Reduction Not Quantitatively Measured The manuscript claims mitigation of hallucinations but provides: No formal hallucination metric No error-type breakdown No calibration analysis Quantitative evaluation (e.g., confidence calibration curves or error taxonomy) would strengthen this claim.

  • 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?

    The paper addresses a clinically meaningful problem with a novel two-stage GSP+MSVR design that effectively formalizes the radiologist’s cognitive workflow into an MLLM reasoning pipeline, supported by convincing ablation results. However, the small dataset (217 cases) lacks statistical significance tests and confidence intervals, full dependence on closed-source commercial APIs undermines reproducibility, and the self-reported “posterior probabilities” are not validated for calibration. Additionally, no direct comparison with human experts is provided, and the core technical contribution is essentially prompt engineering and pipeline assembly without model training, limiting methodological depth. The score could be raised if the authors address statistical rigor, calibration analysis, and model version clarification in the rebuttal.

  • 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.

    Although the authors did not fully address my questions, their work presents some innovations in the application of LLMs to medical imaging.



Author Feedback

We thank the reviewers and AC for their constructive comments. We group concerns into four themes: dataset size, reliability, and generalization; clinical feasibility and reproducibility; baseline strength; and method clarification with failure patterns.

Dataset size, statistical reliability, and generalization (AC, R1-W1, R2-W2, R3-W1). We agree that cohort size and imbalance are important limitations. DCLDs are rare, and collecting confirmed multi-center cases with pixel-level cyst annotations is challenging. To our knowledge, our cohort is among the largest available multi-center datasets for this DCLD differential diagnosis task. Evaluation is not limited to internal validation: Center B is an independent external cohort, and the submitted results show meaningful cross-center performance. The submitted external-center and per-subtype analyses help contextualize reliability under this rare-disease sample size. Per-class analysis (Fig. 3(a) already reports the corresponding results) shows the model is not simply biased toward large classes; PLCH still obtains an F1 of 69.6%, indicating retained discriminative performance even on the minority subtype.

Clinical feasibility and reproducibility (AC, R1-W4, R2-W1, R2-W3, R3). The preprocessing stage is externally validated in the submitted manuscript, supporting the reliability of the upstream segmentation module. The segmentation model can be deployed on a single RTX 4090 GPU, and deployment profiling indicates that the full pipeline is practical for non-emergency DCLD decision support. The system has also been deployed in an anonymized hospital setting and received positive feedback. For reproducibility beyond commercial APIs, the pipeline is modular and can be executed with local open-source MLLM backbones; therefore, commercial MLLMs are high-capability backbones rather than the only executable form. All closed-source models were accessed through OpenRouter, and the detailed model identifiers, inference settings, prompts, preprocessing pipeline, and code are provided in the anonymized repository.

Baseline strength (R1-W6). We understand the concern about the strength of baseline comparisons. We also examined modern supervised architectures, including a Swin Transformer-based encoder and M3T, but they did not outperform the traditional 3D CNN baselines in this rare-disease small-sample setting. This suggests that the core bottleneck is the lack of disease-specific training data rather than the architecture itself. This further supports the rationale of our training-free MLLM reasoning framework: rather than relying solely on fitting a supervised classifier from limited rare-disease data, the proposed framework explicitly combines global lesion distribution with local evidence verification.

Method clarification and failure patterns (R1-W2, R1-W3, R1-W5, R3). GSP is not intended to replace the original CT volume. It provides a global distribution map and slice-index navigation, while MSVR returns to original axial slices for fine-grained morphological verification. This directly addresses the compression-fidelity concern: global spatial information is compressed, but local diagnostic details are checked from the source CT slices. The ablations support this design, as removing GSP causes the largest performance drop and removing MSVR also substantially degrades results. Slice selection is guided by carefully designed prompts; in rare diseases, this training-free design helps avoid overfitting, while learning retrieval becomes promising with larger datasets or data-rich transfers. Failure review showed that the main ambiguity lies between BHD and LIP, which is clinically plausible because both may show heterogeneous cyst size and uneven distribution; distinguishing them requires subtle analysis of cyst shape and local relationships. Thus, remaining errors are concentrated in genuinely difficult differential patterns rather than random majority-class collapse.



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 reviewers are positive about the overall direction of the paper, especially its interpretable reasoning framework and clinically meaningful workflow. The main remaining concerns are about the small and imbalanced dataset, practical feasibility, and the strength of the baseline comparisons. In rebuttal, the authors should focus on clarifying these limitations and keeping the claims aligned with the current level of evidence.

  • 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.

    All reviewers were positive after rebuttal and recognized the clinical relevance of the problem, the interpretable GSP+MSVR reasoning framework, and the meaningful performance gains.



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.

    one accept and two missing after rebuttal.

    while the two missing give WA before rebuttal.

    i read through and agree the paper can be accept subject to minor revision.



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.

    All reviewers gave a weak accept, acknowledging the novel, clinically aligned GSP-MSVR framework, strong interpretability and solid experimental results. The authors reasonably addressed major concerns regarding dataset limitations, baseline selection, practical deployment and method design in their response. While a few inherent drawbacks for rare disease research remain unaddressed, they do not undermine the paper’s core contributions.



back to top