List of Papers Browse by Subject Areas Author List
Abstract
Longitudinal comparison is a routine yet clinically consequential component of chest X-ray (CXR) interpretation, informing disease progression assessment and treatment decisions. Prior temporal CXR models are typically trained as general-purpose pair encoders and adapted with standard progression classifiers, but they do not explicitly capture two key properties of clinical progression assessment: temporal symmetry (for reversible findings, the progression assessment should invert when the temporal order is reversed) and anatomical grounding (decisions rely on localized evidence within specific structures). We propose Bidirectional Anatomy-Aware Progression Perception (BAAP), a lightweight post-training strategy built on top of a pretrained CXR encoder. BAAP encodes temporal symmetry via bidirectional pair augmentation, jointly optimizing each CXR pair in both chronological orders with label inversion to enforce directional consistency and mitigate directional bias. To ground predictions anatomically, we introduce an Anatomy-aware Perception Layer (APL) that aggregates patch-level features within predefined anatomical regions and fuses prior and current region features with their differences for per-region progression classification. Adding only 0.63M parameters (≈0.7% of a ViT-B/16 backbone) and requiring no additional annotations, BAAP achieves 65.90% mean accuracy on MS-CXR-T (+4.78% over the strongest baseline) and 66.4% mean anatomy-level accuracy on Chest ImaGenome Gold (+11.0% over the strongest baseline), demonstrating that a clinically motivated symmetry prior and anatomy-aware design can yield substantial gains for longitudinal CXR progression prediction with minimal overhead. The code is available at https://github.com/chenyuming052/BAAP.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3580_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/chenyuming052/BAAP
Link to the Dataset(s)
mimic-cxr-jpg dataset: https://www.physionet.org/content/mimic-cxr-jpg/2.1.0/
chest imaGenome dataset: https://physionet.org/content/chest-imagenome/1.0.0/
ms-cxr-t dataset: https://physionet.org/content/ms-cxr-t/1.0.0/
BibTex
@InProceedings{CheYum_Bidirectional_MICCAI2026,
author = { Chen, Yuming AND Wang, Hui AND To, Minh-Son AND Zhang, Jianpeng AND Wu, Qi AND Xie, Yutong},
title = { { Bidirectional Anatomy-Aware Post-Training for Longitudinal Chest X-Ray Progression Modeling } },
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 introduces a lightweight post-training framework that explicitly encodes temporal symmetry and anatomical grounding for longitudinal chest X-ray progression modeling, yielding significant performance gains with minimal additional parameters.
- 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 incorporation of temporal symmetry and anatomical localization is well-motivated and aligned with radiological practice. 2: The method achieves strong performance gains with minimal parameter overhead (<0.7%), making it highly deployable. 3: This paper demonstrates robust gains across both disease-level and anatomy-level benchmarks, supported by ablation studies.
- 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:Limits applicability in real-world settings where such annotations may not be available or reliable. 2: The symmetry assumption is heuristically motivated but not rigorously formalized or validated.
- 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.
(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 presents a well-motivated and elegantly designed approach that introduces clinically meaningful inductive biases, achieving substantial performance gains with minimal additional complexity.
- 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
The authors propose a Bidirectional Anatomy-Aware Progression Perception (BAAP) module that incorporates temporal symmetry and anatomical grounding. The method introduces region-level feature aggregation based on anatomical structures, enabling localized reasoning for progression prediction. A lightweight post-training strategy is adopted, adding minimal parameters (<1%) on top of a pretrained encoder while improving 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 proposed method achieves state-of-the-art performance on progression prediction tasks. The paper includes extensive comparisons with both single-image and temporal baselines. The ablation studies are comprehensive, providing evidence for the contribution of individual components. The incorporation of anatomical structure information is well-motivated and clinically relevant, as progression is inherently localized.
- 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 overall idea appears closely related to DiPro (Chen Liu et al., 2025), which also considers bidirectional temporal modeling. However, this work is neither cited nor compared experimentally. It would be important to clarify the differences and include comparisons with DiPro to better assess the novelty of the proposed approach.
2.The model incorporates explicit difference features between current and prior representations. While this is a reasonable design choice for capturing temporal changes, the necessity of explicitly modeling the difference is not fully established. Temporal relationships may already be captured through joint modeling of paired features (zp,zc). Additional ablation studies isolating the contribution of the difference component would strengthen the justification of this design.
3.The proposed method adopts a late fusion strategy, where features are processed independently and combined at the final MLP. This design may limit the model’s ability to capture interactions between temporal and anatomical representations. A discussion or empirical comparison with alternative fusion strategies, such as earlier feature-level fusion, would help clarify the effectiveness of the chosen design.
4.Dependence on anatomical annotations The method relies on predefined anatomical bounding boxes for region-level feature extraction. While the paper claims no additional annotation is required, the approach assumes the availability of anatomical region annotations.
- 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 presents a well-motivated approach with strong empirical results and thorough experimental validation. In particular, the use of anatomical grounding for progression prediction is a meaningful and clinically relevant contribution.
While the proposed method shows strong performance, its novelty compared to prior work such as DiPro appears somewhat incremental. Components such as bidirectional temporal modeling are conceptually similar, and although both approaches consider region-level representations, this work distinguishes itself by incorporating explicit anatomical grounding via predefined regions.
- 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.
My concerns were addressed clearly and convincingly, and I would recommend acceptance. Although I was initially concerned about the dependency on anatomical annotations and the missing citation, the perturbation analysis shows the model’s robustness to annotation uncertainty. The method is also differentiated from DiPro while outperforming it.
Review #3
- Please describe the contribution of the paper
Contributions: 1.The authors propose Bidirectional Progression Training, which accounts for the symmetrical nature of temporal studies and leverages information from prior timestamps to reduce directional bias in longitudinal modeling. 2.The authors propose the Anatomy-aware Perception Layer (APL), which leverages visual features and anatomical cues to construct a rich, anatomy-aware temporal representation, enabling better temporal modeling. 3.The authors highlight that the technique is lightweight, uses a fraction of parameters, and can be integrated with any pretrained CXR encoder. 4.The authors demonstrate state-of-the-art performance on 2 longitudinal CXR datasets, compared with various non-longitudinal and longitudinal techniques.
- 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.
Major Strengths: 1.The anatomy-aware perception layer (APL) is a major strength. It uses ROI pooling over anatomical tokens, along with temporal fusion, to create a representation that helps better model temporal progression. 2.The technique requires training a few parameters and serves as a post-training method for any pretrained CXR encoder.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The authors claim that bidirectional progression training is a major contribution. A prior work (Choi, Doowoong, et al. “A Bidirectional Loss Approach to Imparting Order Sensitivity to Multi-Image Chest X-ray Encoders.” Medical Imaging with Deep Learning-Short Papers. 2025) proposes the same technique of label inversion and training with bidirectional loss. The approach highly overlaps with the proposed bidirectional progression training strategy, and it has not been cited or compared in any of the ablations. Could the authors provide some justification for the same? 2.The authors claim that the technique can be applied to any pretrained CXR image encoder, but their claim lacks ablations. 3.The authors propose temporal difference fusion in the APL module, which concatenates the ROI embeddings from the previous and current studies along with their difference, but the motivation and ablations supporting this design are missing. 4.The authors do not provide an explanation of performance degradation on the pneumonia class with the addition of bidirectional training in Table 3.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.There are limited ablations of the claims made by the authors. 2.The authors have not cited the conflicting work, and they claim label inversion and bidirectional training as a contribution, which has been proposed in prior work as well. (Choi, Doowoong, et al. “A Bidirectional Loss Approach to Imparting Order Sensitivity to Multi-Image Chest X-ray Encoders.” Medical Imaging with Deep Learning-Short Papers. 2025).
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
1.The authors propose to reposition the work by combining three components into a single framework. That said, the current submitted paper, in its current state, highlights bidirectional progression training as a contribution, which, as shared earlier, has already been proposed by authors of “A Bidirectional Loss Approach to Imparting Order Sensitivity to Multi-Image Chest X-ray Encoders.” At this stage, it is difficult to comprehend how the paper will reposition and restructure itself by removing the first contribution. 2.The authors justify the decreased performance of pneumonia, given that the stable class is a minority class, but as stated by the MS-CXR-T description, “improving” is the minority class and not “stable”.
That said, even though I am unable to increase my score, I think the repositioning of the paper prior could have increased the chances of acceptance.
Author Feedback
We thank the AC and reviewers, and are encouraged by recognized strengths: AC notes “strong empirical performance” and a module “attractive to add to an existing encoder”; R1 “well-motivated and elegantly designed” with “robust gains” across benchmarks; R2 “state-of-the-art performance” with “comprehensive” ablations; and R3 calls APL “a major strength”. We address the four AC concerns below, with tags to reviewer points.
(1) Related work and comparison (R2-W1, R3-W1): We will cite Choi et al. (MIDL 2025) and DiPro (NeurIPS 2025), and position BAAP’s contribution as the combination of three elements not jointly addressed by either: temporal symmetry via bidirectional training, anatomy-aware localized progression perception via APL, and a lightweight (0.63M parameters) post-training strategy.
BAAP differs from Choi in three respects. Granularity: Choi predicts whole-image multi-label progression without anatomy-specific regional outputs, whereas ours performs per-region classification over 38 predefined anatomical structures. Supervision: Choi extracts labels from paired reports via LLM, while ours uses 3-class region annotations directly provided by the public dataset without additional preprocessing. Outcome: Choi reports 53.42% mean accuracy on MS-CXR-T, against ours at 65.9%.
BAAP differs from DiPro in three respects. Deployment: BAAP is a CXR-only lightweight post-training module operating on paired CXRs with region-level progression labels, whereas DiPro additionally requires paired EHR time series and demographics. Role of temporal reversal: BAAP inverts 3-class labels under cross-entropy as primary supervision, enforcing directional consistency, whereas DiPro reverses pair order to regularize static-dynamic feature disentanglement. Outcome: under our protocol BAAP reaches 65.9% on MS-CXR-T, against DiPro’s released CXR encoder at 54.03%.
(2) Anatomical-annotation dependency (R1-W1, R2-W4): We acknowledge this reliance on predefined anatomical bounding boxes as a limitation already noted in the submission, with automated anatomy localization specified as future work. For CXR datasets without such annotations, pretrained anatomy detectors like RGRG (CVPR 2023) can supply the required per-region boxes. To assess robustness to imperfect localization, we conducted a perturbation analysis not included in the submission due to space: jittering each bounding box in center and scale by 5% and 10% keeps MS-CXR-T mean accuracy within 0.7 pp of the baseline, showing tolerance to realistic noise.
(3) Difference feature and fusion strategy (R2-W2, R2-W3, R3-W3): Removing only the difference feature while keeping all other components unchanged drops MS-CXR-T mean accuracy from 65.90% to 63.22%, confirming its necessity. We also experimented with cross-attention-based fusion, whose marginal gains did not justify the added parameter cost, so we retained the current design. Late fusion was preferred because progression modeling targets high-level semantic change, whereas early fusion entangles low-level features and inflates computational cost. Broader fusion designs are future work.
(4) Applicability (R3-W2): BAAP is designed as a backbone-agnostic post-training strategy operating on patch-level features from ViT-style encoder. While we avoid overclaiming generality, we validate it on two representative pretrained CXR encoders under identical settings: MS-CXR-T mean accuracy rises from 61.1 to 65.9 on MedST(ICML 2024) and from 57.3 to 64.5 on MGCA (NeurIPS 2022), the latter omitted from the paper under page limits. The consistent gains provide a reasonable empirical basis. Broader backbone evaluation is future work.
(5) Pneumonia class (R3-W4): The pneumonia drop after adding bidirectional training reflects class imbalance in the training data. Pair reversal rebalances improving and worsening but cannot correct the minority stable class, causing a 1.5 pp drop. APL then recovers this loss for a net 5.5 pp gain over baseline.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a lightweight post-training method for longitudinal chest X-ray progression prediction. A Bidirectional Anatomy-Aware Progression Perception module is designed to incorporate temporal symmetry and anatomical localization, to be integrated on top of a pretrained CXR encoder. It is evaluated on two longitudinal CXR datasets with comparisons to single-image and longitudinal baselines and component ablations.
The paper addresses a clinically meaningful problem of progression prediction in radiology. The proposed inductive biases (temporal symmetry and anatomical localization) are well-motivated and aligned with practice. The proposed approach demonstrates a strong empirical performance with substantial gains on longitudinal progression tasks across two datasets (disease- and anatomy-level classification), supported by ablations. The module is lightweight (<1% additional parameters), which makes it attractive to add to an existing encoder.
The following concerns should be addressed: 1) Closely related work should be discussed and cited (e.g., DiPro, Chen Liu et al., 2025; and Choi et al., MIDL 2025 short paper on label inversion/bidirectional loss). An experimental comparison and clearer positioning are needed to substantiate novelty. 2) The approach relies on predefined anatomical region annotations, which may limit applicability if such annotations are unavailable, noisy, or dataset-specific. This assumption should be clarified and discussed as a limitation. 3) Stronger motivation and ablations are recommended for the explicit difference feature in temporal fusion and late-fusion strategies. 4) Additional evidence supporting applicability to “any pretrained CXR encoder,” and an explanation for a reported performance degradation on pneumonia with bidirectional training could help strengthen the claims.
I invite the authors to address reviewers’ concerns in the rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal addressed several of the main concerns including the relationship to prior work, robustness to annotation noise, the temporal difference component, and broader applicability. While some concerns about the novelty framing around bidirectional progression training relative to prior work, as well as class imbalance, remain, I believe that the overall technical merit and strong empirical results place the paper above the acceptance threshold.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper addresses the important problem of longitudinal modeling in chest X-rays and taking this into account in a post-training way. The comments of the reviewers have been adequately addressed. It appears that it is a combination of ideas presented in previous work but still merits presenting to this audience.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors successfully responded to the reviewers’ comments and concerns. The paper presents an interesting methodology with a thorough evaluation.