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

Abstract

Explainable foundation models in radiology are fundamentally limited by the lack of datasets with structured multi-tier annotations (i.e., descriptive prompts, infection/lobar masks, and bounding boxes), which are critical for learning clinically grounded temporal, causal, and diagnostic reasoning. While accurate data realization offers a potential solution, existing data generation approaches either rely on large-scale supervised training or generate raw data without the structured multi-tier annotations. To overcome these limitations, we propose PromptMedCT, the first LLM-guided, training-free disease realization framework that bridges clinical semantics with anatomical constraints to synthesize lung CT scans with multi-tier annotations. The proposed framework can be operated in two modes: a prompt mode for high-fidelity synthesis (as a domain-specific GPT) and a parametric mode for large-scale data generation with multi-tier annotations. Using the proposed PromptMedCT framework (with parametric mode), we construct a dataset comprising 10K 2D CT scans with comprehensive multi-tier annotations. We validate the realism and effectiveness of our framework through a two-step validation strategy that combines comprehensive quantitative and qualitative evaluations. Quantitatively, we perform synthetic-to-real generalization by training a vision–language model on our generated data and testing it on real clinical datasets, achieving 59.22% DICE and 42.46% IoU. For further qualitative evaluation, we perform a medical expert review in which specialists assess a set of generated data and annotations, with 92% of scans rated realistic and 79.5% of the prompts accurately described. Finally, a qualitative comparison against five state-of-the-art domain-specific and general LLM-based generators shows that our method consistently outperforms them in visual realism and multi-tier annotation quality. Additional implementation details: https://github.com/Owais-CodeHub/PromptMedCT

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4691_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/4691_supp.zip

Link to the Code Repository

https://github.com/Owais-CodeHub/PromptMedCT

Link to the Dataset(s)

https://figshare.com/articles/dataset/MedSeg_Covid_Dataset_1/13521488/2 https://zenodo.org/records/3757476

BibTex

@InProceedings{OwaMuh_PromptMedCT_MICCAI2026,
        author = { Owais, Muhammad AND Shafay, Muhammad AND Velayudhan, Divya AND Zubair, Muhammad AND Syed, Naveed AND Werghi, Naoufel AND Rustum, Quatullah AND Hussain, Irfan},
        title = { { PromptMedCT: A Training-Free LLM-Guided Framework for Volumetric CT Data Realization with Quantifiable Disease Manifestations and Multi-Tier Annotations } },
        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

    The paper proposes PromptMedCT, a training-free, LLM-guided framework for generating lung CT data with structured multi-tier annotations, including masks, bounding boxes, and descriptive prompts. The method integrates semantic parsing via LLMs, anatomical template retrieval, morphology-based lesion synthesis, and diffusion-based rendering to bridge clinical language and image generation, enabling controllable and annotation-rich CT data synthesis.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    1. The paper addresses an important and timely problem, namely the lack of structured, multi-tier annotated datasets for training explainable and clinically grounded vision-language models in medical imaging.
    2. The proposed framework is conceptually well designed, integrating LLM-based semantic control with anatomically grounded image synthesis, resulting in a coherent and practical pipeline.
    3. The training-free formulation is particularly appealing, as it avoids reliance on large-scale annotated datasets and improves scalability under privacy constraints.
    4. The ability to generate multi-tier annotations (image, mask, bounding box, and text) within a unified framework is a strong contribution that goes beyond most existing synthesis methods.
    5. The evaluation is comprehensive, including synthetic-to-real generalization experiments and expert-based qualitative assessment, which together provide convincing evidence of realism and usefulness.
  • 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. While the overall framework is well integrated, several components (e. g. , template retrieval, lesion insertion, diffusion-based refinement) rely on existing techniques, and the novelty mainly lies in the system-level combination rather than fundamentally new modeling.
    2. The method is primarily evaluated on 2D CT slices, while the motivation includes volumetric CT generation; the extension to full 3D consistency is not fully explored.
    3. The synthetic-to-real evaluation, although encouraging, remains limited to segmentation performance; additional downstream tasks or broader benchmarks would further strengthen the claims.
    4. The reliance on predefined lesion dictionaries and template databases may limit diversity and realism in more complex or rare pathological scenarios.
  • 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 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.

    (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?

    The paper presents a well-motivated and practically useful framework for generating structured, multi-tier annotated CT data in a training-free manner. The integration of LLM-guided semantic control with anatomically grounded synthesis is both novel and relevant, and the ability to produce diverse annotations within a single pipeline is particularly valuable for advancing vision-language modeling in medical imaging. The experimental validation is reasonably comprehensive and demonstrates promising generalization to real-world tasks. While the technical novelty of individual components is limited and the current evaluation focuses mainly on 2D data and segmentation tasks, the overall contribution is solid and meaningful. I recommend acceptance.

  • 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 introduces PromptMedCT, the first training-free, LLM-guided framework that synthesizes anatomically consistent lung CT scans with structured multi-tier annotations (descriptive prompts, infection/lobar masks, and bounding boxes). The key contributions are: 1.Novel Framework Design: A four-stage pipeline (LLM-guided semantic parsing -> lobar-guided template retrieval -> morphological pathology synthesis -> LLM-guided diffusion) that bridges clinical semantics with anatomical constraints without requiring model training. 2.Dual-Mode Operation: Supports both prompt-mode (high-fidelity, GPT-like interaction) and parametric-mode (scalable batch generation with explicit disease control), offering flexibility for diverse research needs. 3.Large-Scale Annotated Dataset: Construction of a 10K 2D CT scan dataset with comprehensive multi-tier annotations, directly addressing the scarcity of structured medical imaging data. Rigorous Validation: A two-step evaluation strategy combining synthetic-to-real generalization benchmarks (59.22% Dice, 42.46% IoU on real clinical data) and expert-driven qualitative assessment (92% realism rating, 79.5% prompt alignment). 4.State-of-the-Art Comparison: Demonstrates consistent superiority over five contemporary LLM-based and domain-specific generators in visual realism, annotation quality, and generation efficiency.

  • 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 framework’s training-free design is particularly valuable in privacy-constrained medical domains, eliminating the need for large-scale supervised training while maintaining high-quality output. This addresses a critical scalability limitation in existing approaches. 2.PromptMedCT simultaneously produces infection masks, lobar masks, bounding boxes, and radiologist-style reports. This holistic annotation suite enables downstream tasks like vision-language alignment and explainable AI research. 3.The dual quantitative/qualitative evaluation is thorough and well-designed. The synthetic-to-real generalization experiment convincingly demonstrates clinical utility, while expert review provides essential domain validation. 4.The paper is well-organized, with intuitive figures (especially Fig. 1–2) that effectively communicate the framework workflow. The comparison table (Table 1) clearly positions the work within the literature. And the inclusion of a demonstration video and detailed methodological appendices enhances reproducibility and transparency.

  • 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.While the framework is described as applicable to volumetric CT, the presented dataset and experiments focus on 2D slices. A brief discussion of extension to full 3D volumes (e.g., slice-to-slice consistency, volumetric annotation propagation) would strengthen the clinical applicability claim. 2.The current implementation focuses on infection patterns (e.g., COVID-19). While the framework is theoretically extensible, explicit validation on other pathologies (nodules, fibrosis, tumors) would broaden the impact assessment. 3.The reliance on LLaMA3 for prompt parsing introduces potential variability based on prompt phrasing and LLM version. A sensitivity analysis or ablation on prompt robustness would be valuable.

  • 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 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.

    (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?

    PromptMedCT introduces a training-free, LLM-guided framework that unifies structured semantic prompting, explicit parametric control, and complete multi-tier annotation generation within a single controllable pipeline—a genuine advance over existing approaches that either require large-scale supervised training or produce raw images without structured annotations. The four-stage design (LLM-guided semantic parsing, lobar-guided template retrieval, morphological pathology synthesis, and LLM-guided diffusion) is logically coherent and technically sound, enabling the synthesis of anatomically consistent lung CT scans with infection masks, lobar masks, bounding boxes, and radiologist-style reports. The evaluation is comprehensive and convincing: synthetic-to-real generalization experiments demonstrate meaningful cross-domain transfer (59.22% Dice, 42.46% IoU on real clinical benchmarks), while expert review confirms high diagnostic plausibility (92% realism rating, 79.5% prompt alignment). The released 10K annotated dataset addresses a critical resource gap for vision-language model pretraining and explainable AI research. Although the current implementation focuses on 2D slices and infection patterns, these represent natural directions for future extension rather than fundamental limitations.

  • 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 proposes PromptMedCT, an LLM-guided, training-free framework for generating lung CT images together with pixel-level masks, region-level bounding boxes, and text descriptions. The main contribution lies in enabling the generation of structurally consistent synthetic data with multi-tier annotations within a single controllable pipeline, helping to address the scarcity of annotated medical imaging data. However, the novelty primarily stems from system integration rather than fundamental algorithmic advances, and the reliance on LLM priors without task-specific fine-tuning may limit the ability to model complex or atypical anatomical variations. Nevertheless, the construction of a dataset with 10,000 fully annotated CT images, along with a two-stage validation strategy, provides a useful resource and benchmark for the community.

  • 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 training-free framework. The paper presents an LLM-guided approach for CT synthesis that requires no training, offering a practical alternative to conventional generative methods, particularly when data privacy or annotation is a concern. 2.Unified multi-tier annotation. Generating pixel-level masks, bounding boxes, and text descriptions in a single pipeline is a strong contribution, as it directly enables the creation of fully annotated datasets for downstream tasks. 3.Clinically meaningful controllability. The framework allows explicit control over disease attributes such as location, extent, and severity, making the synthetic data more interpretable and clinically relevant. 4.Well-designed evaluation. The combination of synthetic-to-real validation and expert qualitative assessment provides convincing evidence of both realism and practical utility. 5.Useful dataset release. The release of a large-scale dataset with 10,000 samples and comprehensive annotations is a valuable resource for future research.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Strong reliance on LLM priors without domain adaptation, which may limit robustness in modeling complex or atypical anatomical variations. 2.Limited comparison with strong generative baselines, particularly GAN- and diffusion-based methods, making it difficult to fully assess competitiveness.

  • 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.

    (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 proposes a practical and well-motivated framework for training-free CT data synthesis with multi-tier annotations, addressing the important problem of data scarcity. The use of LLM guidance for controllable generation is interesting and supported by reasonable experimental validation.

    However, the method relies heavily on LLM priors without domain-specific adaptation, which may limit robustness, and lacks comparison with strong generative baselines such as GAN- and diffusion-based methods.

    Overall, I lean slightly toward acceptance, depending on the rebuttal.

  • 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

We sincerely thank all the reviewers for their valuable time, careful evaluation, and highly encouraging feedback. We truly appreciate that all reviewers recognized the novelty, practical significance, and potential impact of our training-free, LLM-guided framework for structured CT data realization with unified multi-tier annotations. This work establishes a foundational pipeline for clinically controllable synthetic CT generation and opens a promising direction toward future volumetric 3D CT realization through incremental architectural enhancements and rigorous validation. We are particularly encouraged by the positive recognition of the framework’s ability to bridge clinical semantics with anatomically grounded data synthesis. To further support reproducibility and future research, we will publicly release the generated dataset and detailed algorithmic workflow through a GitHub repository alongside the final version of the paper.



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.

    The reviewers consistently agree that this paper addresses a highly relevant and timely problem in medical imaging, namely the scarcity of structured annotated datasets for vision-language modeling. While the novelty is primarily at the system-integration level and certain aspects such as full 3D validation and broader pathology coverage could be strengthened, these are not fundamental limitations. Given the clear positive consensus, solid technical contribution, and strong practical impact, this paper meets the bar for early acceptance.



back to top