List of Papers Browse by Subject Areas Author List
Abstract
Generating follow-up chest X-rays (CXRs) conditioned on a reference study and progression description can support controllable longitudinal visualization and data completion, and recent autoregressive (AR) models have shown promising fidelity for this task. However, they often display limited lesion-centric spatial awareness, leading to disease-inconsistent follow-up synthesis. In this work, we identify and quantify a previously unrecognized corner/edge attention bias in AR generation, where follow-up queries disproportionately attend to peripheral, non-lesional reference regions. We further introduce LeGend, a lesion-guided framework for longitudinal CXR generation built on Gaussian-Biased Causal Attention (GBCA), a lightweight, causally aligned, plug-and-play correction that injects a lesion-conditioned 2D Gaussian prior into causal self-attention logits to provide sample-specific guidance without modifying the backbone architecture. The prior is obtained from sparse lesion coordinates predicted by an offline vision–language model and smoothed into a disease-relevant salience map to softly steer decoding toward lesion-relevant regions. Experiments show that LeGend improves lesion-relevant attention, image fidelity, and classifier-assessed disease consistency, while yielding more interpretable attention patterns with negligible computational overhead.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3214_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)
N/A
BibTex
@InProceedings{WanYir_LeGend_MICCAI2026,
author = { Wang, Yiran AND Yue, Xiaoyu AND Zhao, Haimei AND Chen, Xinyu AND Zhou, Luping},
title = { { LeGend: Lesion-Guided Longitudinal CXR Generation via Gaussian-Biased Causal Attention } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a lesion-guided framework (LeGend) for longitudinal CXR generation, which aims to generate follow-up images conditioned on a reference image and a progression description. The key contribution is a Gaussian-Biased Causal Attention (GBCA) mechanism that injects a lesion-centered spatial prior into the attention logits of an autoregressive 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.
1.Clear identification of an attention bias issue The paper identifies a non-trivial issue in autoregressive generation, namely the tendency of attention to focus on peripheral regions rather than lesion-relevant areas. This analysis is interesting and provides useful insights into the behavior of such models.
2.Simple and effective design The proposed Gaussian-biased attention is lightweight and easy to integrate into existing autoregressive frameworks without modifying the backbone. This plug-and-play property makes the method practically appealing.
- 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.I don’t really get the motivation of the task I don’t really get the idea of generating follow-up CXRs from a textual progression description, since the text itself is already a high-level summary of the imaging findings. Mapping it back to an image seems inherently ambiguous and ill-posed, as the same description could correspond to many different plausible images. Moreover, this direction does not align well with the typical clinical workflow (image → report rather than report → image), so the practical usefulness of such generation is unclear.
2.The evaluation is not fully aligned with the main claim The paper emphasizes improving lesion-focused attention through the proposed attention mechanism, but the experimental evaluation does not directly validate this claim. Instead, most results are based on global metrics or downstream classification on the whole image. There is a lack of more targeted analysis (e. g. , region-level or lesion-level evaluation) to verify whether the attention is indeed better aligned with clinically relevant regions, making the validation somewhat indirect.
3.Reliance on potentially noisy VLM-derived ROI The lesion regions are obtained from an external VLM, which inevitably introduces localization errors. Since these predicted coordinates are directly used to guide attention during generation, any inaccuracy may be propagated and even amplified in the synthesized images. This raises concerns about the robustness of the method and whether the improvements are limited by the quality of the external VLM rather than the proposed approach itself.
- 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.
(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?
Please ref. to the weakness
- Reviewer confidence
Very confident (4)
- [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.
Only one of my concerns has been directly addressed, namely the motivation. The other two concerns remain largely unanswered.
Regarding the motivation, if the proposed method is intended only for educational visualization, I would not consider this sufficient.
Longitudinal CXR–report pairs are not necessary for learning image–text relationships; single-time-point CXR–report pairs are already sufficient for that purpose. The main value of longitudinal images should lie in modeling or reasoning about temporal changes. However, the current response does not clearly explain why such changes need to be learned from longitudinal images, since many clinically apparent changes can be identified simply by direct comparison between two CXRs.
Review #2
- Please describe the contribution of the paper
This paper addresses follow-up chest X-ray generation conditioned on a reference image and a text description of disease progression. The authors identify a corner/edge attention bias in autoregressive transformers, where attention during follow-up generation focuses on peripheral image regions rather than lesion-relevant areas. They propose LeGend with GBCA, a mechanism that injects a lesion-conditioned 2D Gaussian prior derived from VLM-predicted coordinates as an additive logit bias into self-attention at mid-depth decoder layers, steering generation toward clinically relevant regions without modifying the backbone. Experiments on ICG-CXR show consistent gains in generation quality and downstream disease classification across two AR backbones supporting the method’s generality.
- 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.
Novel technical insight. The corner/edge attention bias seems to be a previously unrecognized phenomenon that is identified and quantified with purpose-built metrics (CAR/EAR/Attn-Overlap). The layer-wise analysis in Fig. 1 and Fig. 4 demonstrates the problem and motivates the proposed solution.
Method. GBCA is lightweight, just an additive bias plus a small MLP scalar. It still yields good improvements (FID drops from 35 to 26). The simplicity and transferability to other architectures strengthen the contribution.
Ablations. The injection depth study (coarse and fine-grained) and the VLM ablation (Qwen vs. Llama vs. random) are well-designed. The random-coordinate control is a intuitive sanity check.
Results. Strong quantitative results. Improvements are consistent across all metrics and substantial, particularly in FID and MS-SSIM.
- 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.Single dataset + no statistical testing. All experiments use ICG-CXR (760 test pairs). The improvements are encouraging but no confidence intervals, significance tests, or multi-run variance is reported. Particularly for AUC and F1 where the differences are modest, how confident are the authors that these gaps reflect genuine improvements rather than noise from a single run? Can the authors comment on the expected variance of their results?
2.VLM dependency underexplored. The method critically relies on VLM-predicted coordinates, yet robustness to prediction quality is only indirectly assessed through the Qwen/Llama comparison. How often does the VLM produce clearly incorrect localizations — wrong side, outside the lung field, hallucinated anatomy? And when it does, does GBCA degrade gracefully or catastrophically? The authors have these predictions already; even a qualitative discussion of failure patterns would help readers assess reliability.
3.Limited clinical evaluation. AUC/F1 from a pretrained classifier is a proxy but the paper uses expressions like “clinical faithfulness” and “clinical fidelity”. In my opinion this sets a higher bar than what the evaluation delivers. How would the authors justify this given no radiologist assessment is included? Do the authors see a path toward validating that the generated images are actually useful in a clinical workflow?
4.Fixed σ with no sensitivity analysis. A single global σ is used for all samples regardless of pathology type. Lung nodules and diffuse effusions have fundamentally different spatial extents. How sensitive is performance to this choice? Would the authors expect the method to work equally well across pathology types, and if so, why?
5.Unclear clinical use case. The paper motivates the task broadly around monitoring disease progression and data completion, but never specifies who would use synthetic follow-ups or how. Is the envisioned application data augmentation for classifier training, clinical decision support, education or something else? The strong downstream classifier results hint at augmentation utility, but this isn’t stated explicitly. Clarifying the intended use case would help readers assess whether the evaluation metrics and the level of clinical validation are appropriate for the claimed application.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper identifies a novel phenomenon, the corner/edge attention bias in autoregressive longitudinal CXR generation. It proposes an elegant, lightweight fix that transfers cleanly across backbones. The ablations are well-designed, and the improvements on generation quality are convincing. My main concerns are around evaluation: no variance or significance testing on a 760-sample test set, no radiologist assessment despite claims of “clinical fidelity”, limited failure analysis of the VLM dependency that underpins the method, and a fixed Gaussian σ with no sensitivity analysis. These gaps are addressable, and the core technical insight is likely to be useful outside this specific task, so I am leaning towards acceptance.
- Reviewer confidence
Somewhat confident (2)
- [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.
The authors have addressed most concerns and I recommend accepting this paper.
Review #3
- Please describe the contribution of the paper
1) The paper introduces a lesion-guided autoregressive (AR) framework LeGend for longitudinal chest X‑ray (CXR) generation. A main contribution is Gaussian-Biased Causal Attention (GBCA), a plug‑and‑play mechanism that injects a lesion‑centric 2D Gaussian prior—derived from sparse lesion coordinates predicted by a vision‑language model (VLM)—directly into the causal attention logits of an AR transformer. The paper identifies a corner/edge attention bias in AR follow‑up CXR generation and shows that GBCA mitigates this issue, improving lesion alignment, image fidelity, and downstream disease‑classification performance. Experiments on ICG‑CXR dataset contains strong improvements over diffusion and AR 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.
1) identification of a systematic peripheral attention bias in AR CXR generation, 2) state‑of‑the‑art performance across FID, CLIP‑T, MS‑SSIM, PSNR, and downstream classification metrics, 3) consistent improvements across both Emu3 and EditAR backbones, supporting generalizability.
- 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) lack of analysis of cases where VLM misses lesions, 2) lack of human evaluation (comparing with radiologist assessment). 3) Evaluations are only on ICG‑CXR, which is derived from MIMIC‑CXR and CheXpertPlus, and may not reflect real‑world longitudinal variability. 4) Lack of discussion on sensitivity to incorrect priors, or over‑dependence on VLM. 5) The current 32×32 grid may limit fine‑grained lesion representation. How does the method scale to higher‑resolution input ? 5) No mention of gpu machines and compute overheads (costs).
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper contains clear contribution, solid experiments, and meaningful clinical relevance, with an interesting identification of peripheral attention bias. However there are several weak spots as outlined above.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers and group responses by motivation, VLM reliability, evaluation, and implementation.
Q1: Task motivation and intended clinical use (R1-W1, R2-W5, Meta). Our task is not text-only report-to-image generation. The model generates a follow-up CXR from a prior CXR and progression text: the prior image provides patient-specific anatomy, while the text specifies disease evolution. The intended use is not direct clinical diagnosis, but controllable longitudinal synthesis for settings with sparse or incomplete follow-up imaging, progression-aware augmentation, robustness testing under clinically meaningful temporal changes, and educational visualization.
Q2: Robustness to VLM prediction (R1-W3, R2-W2, R3-W1/W4, Meta). GBCA uses VLM points as a soft Gaussian logit bias, not hard masks or deterministic editing targets. Generation remains conditioned on the reference CXR, progression text, and autoregressive visual context. Table 3 shows gain with both Qwen and Llama coordinates, despite their localization differences, while random coordinates substantially degrade all metrics, ruling out arbitrary spatial regularization. During dataset construction, we checked VLM reliability: in unilateral cases, ~90% had at least one point on the text-consistent side, and clear wrong-side cases were rare. Grad-CAM from our downstream disease classifier showed overlap between VLM points and high-response disease regions in about half of correctly classified cases. Thus, the prior is coarse but disease-relevant rather than arbitrary noise.
Q3: GBCA design: fixed σ and injection depth (R2-W4, Meta). The Gaussian prior is a coarse salience prior on the 32×32 token grid, not a lesion mask. We use a broad fixed σ to tolerate coordinate noise and avoid overfitting to lesion size without region-level labels, while s(l) adapts the actual bias strength. Table 2 supports mid-layer injection: L14–16 gives the best trade-off, whereas shallow layers lack stable lesion semantics and late or overly wide injection can disrupt refinement.
Q4: Validation about attention improvement (R1-W2) We have included direct attention analysis using Attn-Overlap and CAR/EAR in Fig. 1 and Fig. 4, showing that GBCA shifts attention away from corners/edges and toward the lesion-prior region. Since region-level radiologist labels are unavailable in ICG-CXR, downstream disease classification is used as a proxy for disease-relevant consistency, not as definitive clinical validation.
Q5: Statistical reliability and single-run variance (R2-W1) Our conclusions are based on full test-set evaluation rather than selected examples. GBCA consistently improves AUC/F1, substantially outperforms random-coordinate guidance, and remains effective across different VLMs and two AR backbones. Table 1 reports three-seed averages; the observed variance was small (below 0.5% for AUC/F1), supporting the stability of the improvements.
Q6: Clinical claims and evaluation scope (R2-W3, R3-W2/W3) We will replace “clinical fidelity” with “classifier-assessed clinical consistency” and clarify that deployment requires radiologist reader studies and prospective evaluation. Evaluation is limited to ICG-CXR, which to our knowledge is the only public benchmark for text-controlled longitudinal CXR synthesis, following prior work ProgEmu (MICCAI2025). ICG-CXR is built from hospital-collected MIMIC-CXR and CheXpertPlus radiographs and captures substantial real-world longitudinal progression variability. Broader external validation remains future work.
Q7: Resolution and compute overhead (R3-W5/Meta)
GBCA is grid-agnostic: for a denser H×W token grid, VLM coordinates can be remapped and σ scaled with grid size. Higher-resolution tokenization may better capture small lesions. The dominant cost remains base AR attention; GBCA only adds a Gaussian bias to existing logits and no extra generative network. We fine-tuned with LoRA for 6k steps on 2×RTX A6000 GPUs.
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 rebuttal should address the reliance on VLM-predicted lesion coordinates, including how the method behaves when localizations are inaccurate, missing, or hallucinated. The authors should also clarify the motivation and intended clinical use case of the task, since generating follow-up CXRs from textual progression descriptions is seen as ambiguous and not aligned with the typical image-to-report workflow. The fixed Gaussian σ across all pathology types and the choice of injection depth need to be clarified. Finally, scalability to higher-resolution inputs and compute overhead should be briefly discussed.
- 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 presents an interesting methodological contribution and is sufficiently evaluated. The rebuttal sufficiently addressed most points.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The post-rebuttal opinions are mixed. Reviewer #1 maintains rejection, arguing that only the motivation was partially addressed and that the necessity of learning from longitudinal CXR–report pairs remains insufficiently justified. Reviewer #2 recommends acceptance, stating that most concerns were addressed. Reviewer #3 does not provide a visible post-rebuttal decision, but the initial review was weakly positive.
Overall, the paper remains borderline but leans slightly positive. The proposed GBCA mechanism is simple, technically interesting, and addresses a meaningful attention-bias issue in autoregressive follow-up CXR generation. However, important limitations remain, including unclear clinical use, reliance on VLM-predicted lesion coordinates, lack of radiologist evaluation, single-dataset validation, and limited evidence that longitudinal generation is necessary beyond direct image comparison. The authors should provide the full open-source implementation to improve the reproducibility.
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 clinical use case and reliance on VLM-derived lesion coordinates remain limitations, but the paper identifies an interesting peripheral attention bias and proposes a simple, well-motivated GBCA mechanism with solid ablation evidence.