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

Abstract

Text-conditioned diffusion models have enabled chest X-ray (CXR) synthesis, yet when prompted with multi-finding radiology reports they frequently omit findings or place them at anatomically incorrect locations. Existing approaches improve this through reinforcement learning or adapter training, but their reliance on additional training limits portability across base models. We propose AMFG (Anatomy-grounded Multi-Finding Guidance), a training-free framework applied at inference time to a frozen CXR diffusion model. AMFG comprises three complementary guidance signals: (1) finding presence guidance using a pretrained pathology classifier, (2) anatomical grounding guidance that constrains cross-attention maps to clinically correct regions via segmentation masks, and (3) finding interaction guidance that enforces spatial consistency among co-occurring findings. All components use only existing public models with no additional training. Experiments on MIMIC-CXR demonstrate that AMFG achieves the highest anatomical correctness among all compared methods and that this accuracy translates into superior downstream classifier performance, demonstrating that explicit anatomical grounding—rather than additional training alone—is key to clinically useful synthetic CXR data.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4880_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

MIMIC-CXR: https://physionet.org/content/mimic-cxr/

BibTex

@InProceedings{HanYeo_AMFG_MICCAI2026,
        author = { Han, Yeon Gyu AND Jung, Junah AND Park, Chang Min AND Lee, Dongheon},
        title = { { AMFG: Anatomy-Grounded Multi-finding Guidance for Training-Free Chest X-Ray Generation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper presents AMFG, a training-free guidance framework for generating multi-finding chest X-ray images at inference time using only public models. The main contributions include: (1) Proposing a training-free framework that steers a frozen CXR diffusion model to generate anatomically faithful multi-finding CXR images at inference time, using only publicly available pretrained models without additional training. (2) Demonstrating through systematic ablation that the bottleneck in multi-finding CXR image generation is anatomical spatial supervision rather than model capacity. (3) Providing empirical evidence that anatomically correct synthetic CXR data directly improves downstream classifier performance with the largest gains on rare findings, suggesting that spatial accuracy determines clinical utility.

  • 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.AMFG guides a frozen diffusion model at inference using only publicly available pretrained models via three complementary signals. It requires no fine‑tuning or adapter training, making it easy to deploy across different base models.

    2.AMFG directly addresses poor spatial calibration in CXR diffusion models by constraining cross‑attention to clinically correct regions via segmentation masks.

    3.Comprehensive experiments evaluate AMFG’s generation quality through multiple complementary metrics (FID for image quality, mAUROC for finding presence, Dice and MF‑Acc for spatial localization), and demonstrate downstream clinical utility via classifier augmentation, collectively proving that AMFG improves both finding consistency and anatomical 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.

    1.In latent diffusion models, x typically denotes the image, and z=enc(x) the VAE-compressed latent, with the diffusion process operating on z. While Fig. 1 and Eq. (6) use this convention correctly, Section 3.1 contains errors: it incorrectly mixes the predicted clean image x0 and zt through reversed diffusion formula​. Additionally, Section 3.2 and 3.3 use dec(x0), which is redundant since x0 should already be in pixel space. These inconsistencies may confuse readers and should be corrected.

    2.Fig. 2 compares generated CXRs across methods, but no highlights (e. g. , arrows, bounding boxes, or overlaid masks) indicate where findings are correctly or incorrectly located. And it does not indicate what the baseline method is. Adding such annotations would better support the claimed advantages of AMFG.

  • 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 addresses a critical problem in multi‑finding CXR image generation with a novel training‑free guidance framework. The method requires no fine‑tuning or adapter training, making it portable across base models. Comprehensive evaluation convincingly demonstrates generation quality and clinical utility. Minor weaknesses exist in presentation (e.g., notations, figure) but do not undermine the core contributions. Overall, the work is well‑structured, technically sound, and offers practical value for medical image synthesis.

  • 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 proposes AMFG, a training-free inference-time guidance framework for steering a frozen chest X-ray (CXR) diffusion model to generate multi-finding images with anatomically correct placement. The method combines three complementary guidance signals: (1) a finding presence loss using a pretrained pathology classifier (CheXNet/TorchXRayVision), (2) an anatomical grounding loss that constrains cross-attention maps to clinically appropriate regions using a pretrained segmentation model (CheXmask) and a fixed finding-to-anatomy mapping, and (3) a finding interaction loss that penalizes attention overlap between distinct findings and enforces bilateral symmetry for bilateral findings. Experiments on MIMIC-CXR show that AMFG achieves superior anatomical correctness (Dice, MF-Acc) compared to both training-free and training-based baselines, and that this translates into improved downstream classifier performance.

  • 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 core insight is compelling and clearly articulated: faithful multi-finding CXR generation requires not just ensuring findings are present, but controlling where they appear and how they interact spatially. This is a real limitation of existing methods that the paper addresses head-on.

    • The training-free design is a genuine practical advantage. By operating entirely at inference time on a frozen model, AMFG is portable across base models, which is a significant benefit over methods like XReal that require retraining adapters for each new model. This makes the approach broadly applicable.

    • The three-component guidance framework (what/where/how) is elegant and well-motivated. Each component has a clear purpose, and the ablation study (Table 2) convincingly shows the contribution of each, with the anatomical grounding loss (L_attn) being the dominant driver. This is a clean and interpretable result.

    • The evaluation is thoughtful. The authors introduce anatomical correctness metrics (Dice, MF-Acc) that go beyond standard FID/mAUROC, and they validate these metrics with cross-validation using an independent segmentor (PSPNet, r=0.94). The downstream classification experiment (Table 3) is particularly valuable because it connects generation quality to actual clinical utility.

    • The paper is well-written, concise, and clearly structured. The method section is easy to follow, and the figures effectively illustrate both the approach and the results.

  • 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 Dice evaluation metric shares the same segmentation model (CheXmask) used during guidance. Although the authors address this with a PSPNet cross-validation (r=0.94), this is a correlation check, not an independence guarantee. A stronger validation would be to evaluate Dice entirely with a segmentor that was never used during generation. The current setup risks inflating Dice scores even if the cross-validation suggests otherwise.

    • The finding-to-anatomy mapping (phi) is fixed and manually defined. The paper only demonstrates mappings for a limited set of findings, such as cardiac findings to heart and pulmonary findings to ipsilateral lung. What about findings that span multiple regions or don’t have a clear anatomical anchor (e.g., support devices, fractures)? The authors mention that such findings are guided by L_cls only, but this means the method effectively falls back to ungrounded guidance for a non-trivial subset of CXR findings.

    • The method is evaluated on a single base model (RoentGen) and a single dataset (MIMIC-CXR). The portability claim is a central selling point, but it remains unvalidated. Demonstrating AMFG on even one additional base model (e.g., a different CXR diffusion model) would substantially strengthen the paper.

    • The downstream classification experiment (Table 3) uses only 1,000 synthetic images for augmentation. It would be informative to see how performance scales with the number of synthetic images, and whether there is a saturation point or diminishing returns.

    • The inference time overhead (1.3x) is reported but not broken down by component. How much of this is from the segmentor, the classifier, and the gradient computation? This matters for practical deployment.

  • 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
    • The noise-level-aware guidance schedule g(t) = sqrt(1 - alpha_bar_t) is a nice detail. The ablation showing it outperforms uniform guidance validates this design choice.
    • The bilateral symmetry loss (L_sym) is an interesting idea, but its impact seems modest in the ablation. Could the authors comment on how frequently bilateral findings appear in the test set?
    • Minor: The paper title is suppressed in headers (“Title Suppressed Due to Excessive Length”), which should be fixed for the camera-ready.
    • The qualitative examples in Figure 2 are helpful but limited to 4 cases. More examples (perhaps in supplementary) would be welcome.
    • How does AMFG handle conflicting guidance signals? For instance, if L_cls pushes the model toward a finding but L_attn constrains it to a region where the diffusion model struggles to render that finding?
  • 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?

    This paper makes a clear and well-executed contribution to medical image synthesis. The insight that anatomical grounding, not additional training, is the key bottleneck for multi-finding CXR generation is convincingly demonstrated through both ablation and downstream evaluation. The training-free design offers genuine practical advantages. The evaluation is thorough, with appropriate metrics and a downstream task that validates clinical utility. The main limitations (shared segmentor between guidance and evaluation, single base model, fixed anatomy mapping) are real but do not undermine the core contribution. The method is simple, principled, and effective, which is exactly what makes for a good contribution.

  • 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 work addresses the gap of text-conditioned diffusion models for Chest X-Ray (CXR) image synthesis. The text conditions are provided by reports and these findings are usually more than one for each CXR report. The work focuses on the improvement of multi-finding text-conditioning that suffers from anatomically incorrect locations and the omission of findings present in the report. The training-free approach checks the presence of the finding and constrains it to an anatomical location. For example, pleural effusion can only exist in the lungs. Using a publicly available CXR synthesis model, they proved the free-training approach improved downstream pathological classification.

  • 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 methodology’s strength lies in the use of publicly available data and models to prove that a simple idea such as anatomical spatial constrains can improve models for other tasks, such as classification. The explanation of the architecture and used components for the training-free approach is solid and understandable and the experimental campaign includes varied approaches from other authors.

  • 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 non-overlap condition of the attention maps of the findings “must not excessively overlap”. This criteria is quite vague and I think it can change depending on the finding, so it needs further explanation. The evaluation metrics of Dice score brings up some doubts about the information leakage that the attention mask in the inference gives to the model, that it’s input later on is used for the classifier and it’s output GradCAM. Although the weights of these two models are not shared.

  • 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 work is well written and comprehensible, it seems to be reproducible due to it’s detailed description. And for the most part, the evaluation process seems robust. From the novelty point-of-view, the idea to condition text using spatial anatomical masks is interesting but can be seen in other fields as well. It is not a groundbreaking work but I would say it would an accept.

  • 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 thank the reviewers for their positive and constructive feedback. We clarify the main remaining concerns below.

(1) Central contribution and novelty. We agree that individual ingredients such as classifier guidance, attention guidance, and anatomical masks have precedents. AMFG’s contribution is the clinically grounded integration of these ingredients for the what/where/how failure modes in multi-finding CXR generation without additional training. The submitted ablation shows that the classifier guidance improves finding presence but the anatomical grounding guidance drives the localization gain, and Table 3 connects this anatomical correctness to downstream utility, with the largest improvement on rare findings. Thus, the key message is that anatomical grounding, not visual realism or additional training alone, is critical for clinically useful synthetic CXR data. Consistently, AMFG outperforms the trained XReal adapter on anatomical metrics (Dice .458 vs .421; MF-Acc 42.6% vs 38.4%).

(2) Dice evaluation and possible leakage. We agree that using CheXmask as the anatomical ROI source in both guidance and Dice evaluation is a limitation. However, the evaluated localization signal is not the guidance loss itself: Dice uses GradCAMs from an independent CheXpert-pretrained evaluation classifier, different from the NIH ChestX-ray14 classifier used for the classifier guidance. The PSPNet cross-validation reported in Sec. 4.2 (r=0.94, p<0.001) supports robustness to the mask source, although it is not a full independence guarantee. Importantly, the downstream classification experiment in Table 3 does not use segmentation masks at evaluation time and still shows AMFG’s largest gain, providing additional evidence that the improvement is not solely an artifact of the shared ROI source.

(3) Scope of the finding-anatomy mapping and the overlap regularizer. The mapping is intentionally conservative and transparent. AMFG grounds findings only when a reliable anatomical anchor is available; findings without a single target region are guided by the classifier guidance only, as stated in Sec. 3.3.We therefore do not claim universal anatomical grounding for every CXR label; expanding the mapping via radiology knowledge graphs remains future work. Likewise, the overlap regularizer is not a hard clinical assumption that pathologies cannot co-exist or overlap in the image. It is a soft attention-space regularizer that discourages multiple finding tokens from collapsing onto the same region. The ablation also shows that the anatomical grounding guidance is the primary driver, while the finding interaction guidance provides a smaller refinement. We will clarify this wording in the camera-ready.

(4) Empirical scope and presentation fixes. We agree that empirical validation currently uses one base model and one dataset, as acknowledged in Sec. 5.Our portability claim refers to AMFG being weight-preserving, requiring no base-model or adapter training, and using standard LDM interfaces, rather than completed empirical validation across CXR diffusion backbones; cross-architecture empirical validation remains future work. Systematic scaling of synthetic augmentation and per-component overhead profiling are also valuable follow-up directions. In the camera-ready, we will correct latent/image notation, running header, Figure 2 caption wording, and typographical issues. These are presentation and clarification fixes only and do not change the method, datasets, experimental protocol, evaluation protocol, or reported results.



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.

    all reviewers agree the experiments are well thought and effective, the paper is clearly written. The remaining concerns can be addressed in rebuttal or camera-ready before acceptance



back to top