List of Papers Browse by Subject Areas Author List
Abstract
Dense anatomical correspondence across computed tomography (CT) scans is essential for longitudinal disease analysis, inter-phase alignment, and quantitative assessment. However, establishing reliable correspondence in longitudinal CT remains challenging: scan pairs are relatively scarce and temporally consistent annotations are costly to obtain, limiting the supervision for learning-based methods. Moreover, substantial appearance variations caused by respiration, pathology, and acquisition differences further challenge explicit registration and hinder generalization. In this work, we propose a diffusion-guided anatomical position encoding framework that leverages large-scale cross-subject CT data to learn generalized anatomical consistency and transfer it to intra-patient longitudinal tracking. Intermediate features from a pretrained CT generative diffusion model are leveraged as anatomical priors to supervise voxel-wise position embeddings, enabling correspondence learning without spatial alignment or task-specific annotations. To promote globally consistent and anatomically discriminative matching, we introduce dual momentum contrastive dictionaries for body-wide feature regularization, followed by a coarse-to-fine refinement network for sub-voxel correspondence precision. We evaluate the proposed method on lung landmark tracking in the DIR-Lab 4DCT dataset and multi-organ lesion tracking in the Deep Longitudinal Study (DLS) dataset. On DIR-Lab, our method achieves the lowest average TRE of 1.44 mm; on DLS, it attains a MED of 5.7 mm and a CPM@10mm of 87.63\%, outperforming state-of-the-art unsupervised methods and approaching supervised trackers without using lesion annotations during training. These results highlight the potential of diffusion-guided anatomical priors for accurate and annotation-efficient longitudinal CT correspondence. The code is available at: https://github.com/DlutMedimgGroup/diffusion-ct-tracking.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0785_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/DlutMedimgGroup/diffusion-ct-tracking
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhuMin_DiffusionGuided_MICCAI2026,
author = { Zhuang, Mingrui AND Zhang, SongYao AND Zhang, Qinhe AND Xu, Rui AND Liu, Ailian AND Fan, Xin AND Wang, Hongkai},
title = { { Diffusion-Guided Anatomical Position Encoding for Dense Longitudinal CT Correspondence } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper’s main contribution is a framework that repurposes a pretrained CT generative diffusion model (MAISI) as a source of anatomical knowledge to learn dense voxel-level position embeddings for longitudinal CT correspondence. the paper shows that generative diffusion models, originally designed for image synthesis, can be effectively “recycled” as anatomical priors for correspondence learning, bridging the gap between large-scale cross-subject pretraining and intra-patient longitudinal tracking.
The key insight is that a diffusion model trained to synthesize realistic CT images must internally encode consistent anatomical structure across subjects. Rather than using these diffusion features directly for matching (which yields coarse results, as shown in their Fig. 3), the authors distill the similarity structure of the diffusion features into a compact embedding network. This is done by aligning similarity distributions between the diffusion teacher and embedding student via KL divergence. so if two voxels look anatomically similar to the diffusion model, they should also be similar in the learned embedding space.
- 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 idea of reinterpreting a generative CT diffusion model as a carrier of cross-subject anatomical consistency and distilling its similarity structure (rather than raw features) is insightful and well-executed.
2.Strong quantitative results without task-specific annotations.
3.Table 3 (page 8) clearly isolates the contribution of each component. it confirms the value of diffusion-based supervision beyond standard contrastive learning.
- 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.UAE* is a direct extension of SAM that originated this line of work, with reported improvements on the DLS benchmark. Its omission from Table 2 weakens the experimental validation.
- Bai, et al. “UAE: Universal anatomical embedding on multi-modality medical images. “ Medical Image Analysis 103 (2025): 103562.
- The method relies entirely on MAISI’s learned distribution. The paper does not discuss how the method handles new/disappeared lesions in the DLS task, which is clinically important. Does the method always predict a match?
3.the adapter architecture, search window size for Stage 2 refinement, and dictionary sampling strategy lack detail.
4.The concepts of finding dense correspondence by using coarse-to-fine estimation, local search windows, attention-weighted localization, and self-sup learning have been studied in the optical flow literature (e. g. , Contrastive Random Walk*). This line of research should also be discussed.
- Bian, et al. “Learning pixel trajectories with multiscale contrastive random walks. “ Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a well-motivated idea: using diffusion model features as structural priors for anatomical correspondence. Solid empirical results on two benchmarks. However, the missing UAE baseline, limited failure case discussion, and inadequate coverage of related work represent gaps that prevent a stronger recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 committed to adding a discussion of the missing related work. I find the repurposing of a generative model for longitudinal correspondence matching to be an interesting idea and a reasonable contribution in its own right. I maintain my previous decision to accept.
Review #2
- Please describe the contribution of the paper
This paper proposes a diffusion guided framework for dense anatomical correspondence in longitudinal CT. A voxel wise embedding network is trained using intermediate features from a pretrained diffusion model as anatomical priors, combined with contrastive learning for global discrimination and a coarse-to-fine refinement stage for matching. The method operates without registration or correspondence annotations and is evaluated on lung landmark tracking and longitudinal lesion tracking.
- 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 problem addressed in the paper is clinically relevant. Longitudinal correspondence in CT is important for disease progression assessment and remains challenging due to appearance variability. 2.The method avoids correspondence supervision and does not require explicit registration, which is valuable given the limited availability of annotated longitudinal data. 3.The method achieves competitive performance and approaches supervised methods on the DLS benchmark despite being trained in an unsupervised manner. 4.The ablation study provides a clear breakdown of the contributions of diffusion guidance, contrastive learning, and the refinement stage, showing their complementary roles.
- 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 contribution is somewhat incremental relative to prior work. While using diffusion features as supervision is interesting, prior work has already shown that such features can be used for correspondence and the proposed formulation can be seen as an extension that distills them into a task-specific embedding space. 2.The coarse-to-fine design and local refinement network are relatively standard and contribute significantly to the overall performance, which may limit the perceived novelty of the method. 3.The contrastive formulation assumes that features sampled from the momentum dictionary act as negatives, however, anatomically similar regions may be incorrectly treated as negatives. The impact of such false negatives is not discussed. 4.The role of diffusion features is not deeply analyzed. While they are assumed to encode anatomical structure, there is limited investigation of which layers or timesteps are most informative or how sensitive the method is to these choices.
- 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?
This work is solid and clinically relevant with competitive performance. But the novelty is somewhat incremental and the contribution of the diffusion component is not fully isolated from the refinement stage.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 addressed my concerns sufficiently.
Review #3
- Please describe the contribution of the paper
This paper proposes a diffusion-guided anatomical position encoding framework for dense longitudinal CT correspondence. The key idea is to leverage intermediate representations from a pretrained generative diffusion model as anatomical priors, and distill them into voxel-wise embeddings for correspondence learning without requiring spatial alignment or task-specific annotations.
- 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.
Instead of using diffusion for generation, this paper leverages intermediate denoising features as structural priors. A dual momentum dictionary mechanism is designed to enforce global anatomical consistency and discriminability in the learned embedding space, improving robustness across different anatomical regions.
- 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 paper assumes that intermediate diffusion features encode anatomically consistent structural priors, which is a key foundation of the proposed method. However, this claim is only loosely motivated and lacks sufficient empirical or theoretical justification. In addition, the choice of extracting features from the first and third decoder blocks of the diffusion U-Net is not well justified, and it remains unclear whether other layer selections or multi-scale combinations would yield different results. Furthermore, the method assumes that spatially corresponding coordinates within overlapping patches can serve as positive pairs. Due to anatomical deformation in longitudinal CT, this assumption may not strictly hold, especially in regions with large motion. Since Stage 2 relies on coarse correspondences predicted by Stage 1, inaccurate coarse matching may lead to unreliable positive pair construction, potentially introducing noisy supervision and raising concerns about the robustness and stability of the overall framework.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
The paper presents an interesting attempt to leverage pretrained diffusion models as anatomical priors for dense longitudinal CT correspondence. The overall framework is well-designed, with a reasonable coarse-to-fine strategy that combines global anatomical consistency and local refinement, and achieves competitive results among unsupervised methods. However, several key design choices are insufficiently justified. In particular, the assumption that diffusion features inherently encode anatomically consistent priors lacks empirical validation, and the selection of specific diffusion layers is not well motivated. In addition, the construction of positive pairs in Stage 2 relies on spatial alignment within overlapping patches, which may not hold under anatomical deformation. This issue is further compounded by the dependency on Stage 1 coarse matching, raising concerns about error propagation and the robustness of the training process. Overall, while the method is promising, the lack of deeper analysis and justification for these core assumptions weakens the technical contribution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 overall framework is well-designed and achieves competitive results among unsupervised methods.
Author Feedback
We thank the reviewers and meta-reviewer (MR) for recognizing the clinical relevance, strong results, and clear ablations.
1.Diffusion priors and layer choices (R3, R4, MR). We thank the reviewers for raising this important point. Our motivation for using diffusion features as anatomical priors is twofold: (1) prior work shows that intermediate features of latent diffusion models support cross-image semantic correspondence, with semantically matched regions producing similar intermediate representations (DIFT; Diffusion Hyperfeatures); and (2) to synthesize a realistic CT volume, a diffusion model must identify the anatomical location of each voxel and recover the texture appropriate to that location, making its intermediate features naturally informative for cross-scan anatomical correspondence. We also observed this in our preliminary analysis: in Fig. 3 (third column), diffusion-feature matching is anatomically coherent despite being spatially coarse. For layer choice, we determined it empirically: the first and third decoder blocks gave the best correspondence accuracy, likely because they balance global anatomical context and local detail. We will clarify this rationale, selection procedure, and Fig. 3 interpretation.
2.Relation to UAE (R2, MR). We thank the reviewers for highlighting UAE. We will add it to related work and clarify that UAE-S is not directly comparable as a fully annotation-free baseline here, because it uses TotalSegmentator-derived organ masks to train an additional semantic branch, while TotalSegmentator itself relies on extensive segmentation annotations. On the same benchmark, UAE-S reports 84.06 CPM@10mm and 5.4 MED, versus 87.63 and 5.7 for ours. Thus, our method achieves higher CPM with comparable MED, without organ-mask supervision.
3.Pair construction and coarse-to-fine refinement (R3, R4, MR). We agree that overlap-based training, as used in SAM/UAE, is not fully aligned with test-time longitudinal CT matching. Our coarse-to-fine design mitigates this mismatch: Stage 1 uses diffusion-feature correspondence to support global coarse matching beyond overlap alone, while Stage 2 refines local texture around a coarse match, where the mismatch is much smaller. Thus, compared with prior overlap-based methods, our unsupervised training better reduces the train-test gap. Regarding negatives, computational cost creates a trade-off between filtered negatives with a small pool and unfiltered negatives with a large pool. Under body-wide sampling, such false negatives are rare and have negligible impact on the overall contrastive objective. Our “w/o Dual Momentum Dictionaries” ablation corresponds to the former setting, and the results show that the larger negative pool is more effective. We will clarify this in the revision.
4.Failure modes and limitations (R2, R4, MR). We thank the reviewers for raising this important clinical concern. Our method remains applicable as long as the underlying anatomy is largely preserved, even if small lesions appear, disappear, shrink, or grow, because it matches anatomical locations rather than lesion appearance. It may fail when the anatomy is substantially altered, such as after resection of a large space-occupying lesion, where one-to-one correspondence becomes inherently ambiguous. We will clarify this task boundary and failure modes.
5.Other issues (R2, R3). We will improve reproducibility by adding technical details of the adapter, local search window, and momentum-dictionary sampling/update strategy. We also agree that coarse-to-fine matching and local search share key ideas with prior optical-flow and self-supervised correspondence methods, such as Contrastive Random Walks, and will discuss this more clearly. We will clarify that our novelty lies not in coarse-to-fine matching itself, but in using CT diffusion priors to learn annotation-free anatomical position embeddings; the submitted ablation already disentangles diffusion guidance from refinement.
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.
This paper proposes a diffusion-guided framework for dense longitudinal CT correspondence. It leverages intermediate representations from a pretrained generative diffusion model as anatomical priors and distills their similarity structure into voxel-wise embeddings without correspondence or registration supervision.
The reviewers agree that the idea is interesting and the topic clinically relevant. The method demonstrated strong performance on longitudinal landmark and lesion tracking benchmarks and is supported by clear ablations.
However, several concerns have been raised. These are mainly conceptual and presentational, rather than fundamental flaws and they appear to be addressable in a rebuttal. The key issues to address are
- The key assumptions should be clarified and justified. Specifically, whether the intermediate fusion features capture anatomically meaningful structures, and why these particular diffusion layers were selected.
- Discuss missing or closely related baselines (e.g., UAE).
- Clarify assumptions behind positive/negative pair construction and potential error propagation from coarse to fine stages
- What are failure modes and limitations? How does the method handle lesion appearance/disappearance?
- 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 appropriately addressed the concerns raised by the reviewers in the rebuttal and all reviewers now agree that this paper should be accepted.
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.
This paper received consistently positive evaluations following the rebuttal. Reviewer 4 upgraded the rating from Weak Reject to Accept, expressing support for the proposed paradigm of using diffusion features as priors. Reviewers 2 and 3 maintained their Accept recommendations, while noting that refinements to the related-work discussion and additional ablation analyses would further strengthen the paper and should be incorporated in the camera-ready version. Overall, the framework is well designed and demonstrates competitive performance among unsupervised methods.
In light of the post-rebuttal discussion and the overall reviewer consensus, I recommend Accept. For the camera-ready version, the authors are encouraged to further strengthen the paper by expanding the discussion of relevant prior work and incorporating the reviewers’ additional suggestions where appropriate.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers have agreed on the acceptance of the paper. This work is solid and clinically relevant with competitive performance. The repurposing of a generative model for longitudinal correspondence matching to be an interesting idea.