List of Papers Browse by Subject Areas Author List
Abstract
Accurate automated tumor segmentation in PET-CT imaging is critical for clinical diagnosis and radiotherapy planning. However, limited annotated data, the high computational cost of volumetric modeling, and intrinsic physical discrepancies between metabolic PET and anatomical CT modalities pose substantial challenges. To address these issues, we propose CDFP-Net, a dual-stream framework that integrates generative diffusion priors with cross-modal dynamic graph fusion. Specifically, we employ a frozen, self-supervised pre-trained Denoising Diffusion Probabilistic Model (DDPM) as a robust feature prior, augmented with lightweight adapters to enhance modality-specific representations under limited supervision. To effectively bridge cross-modal discrepancies, we design a modality-asymmetric dynamic graph fusion mechanism in which metabolically active PET regions act as spatial anchors to query and aggregate complementary anatomical boundary cues from CT. To alleviate the computational burden of full 3D processing while preserving global semantic topology, we further introduce a Multi-Angle Maximum Intensity Projection (MA-MIP) strategy for efficient volumetric context modeling. Extensive experiments on three multi-center datasets (AutoPET, HECKTOR, and PCLT20K) demonstrate that CDFP-Net consistently outperforms state-of-the-art methods in both accuracy and robustness, highlighting its strong potential for precise biological target volume (BTV) delineation in clinical practice. Source code is available at https://github.com/WeiMinqin/CDFP-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2844_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/WeiMinqin/CDFP-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WeiMin_CDFPNet_MICCAI2026,
author = { Wei, Minqin AND Wu, Shangqian AND Xu, Chang AND Qian, Yurong AND Tang, Yongjun AND Song, Jinmiao AND Deng, Lei},
title = { { CDFP-Net: Cross-Modal Dynamic Fusion with Diffusion Priors for PET-CT Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors present CDFP-Net, a 2.5D dual-modality framework for lesion segmentation in PET-CT images. The framework leverages several state-of-the-art techniques in combination, to achieve good segmentation scores. First, the 2.5D approach is achieved through a variant of the MIP approach based on generating several projections along multiple fields of view. Then, during the pre-training phase, a generative model is trained to generate the 2.5D representations. During the task training phase, the encoder part of the diffusion model is frozen while a cross-modality fusion model based on graph convolutional networks is trained with a mix of a novel “spatially standardized hard-mining loss” and focal loss.
- 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 paper’s main strength lies in the effective combination of existing state-of-the-art components — including the diffusion pre-training approach from [1], squeeze-excitation modules from [2], and the Cross-Modality Heterogeneous Graph Fusion (CMHGF) from [3] — into a coherent and functional framework that achieves good overall performance. The ablation studies are also noteworthy, particularly the comparison between the Q=PET and Q=CT settings.
[1] Purma, V. , Srinath, S. , Srirangarajan, S. , et al. : Genselfdiff-his: generative selfsupervision using diffusion for histopathological image segmentation. IEEE Transactions on Medical Imaging 44(2), 618–631 (2024) [2] J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, pp. 7132–7141, 2018. [3] Zheng, S. , Ye, X. , Yang, C. , et al. : Asymmetric adaptive heterogeneous network for multi-modality medical image segmentation. IEEE Transactions on Medical Imaging 44(4), 1836–1852 (2025)
- 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 has two major weaknesses. First, the state-of-the-art comparison is misleading: the baselines reported are vanilla implementations (e. g. , basic nnUNet), whereas published methods in the literature substantially outperform the values reported — for instance, the best participants in the AutoPET challenge achieved mean Dice scores between 0.74 and 0.79 [1], compared to the authors’ result of 0.63; similarly for HECKTOR, SE-nnUNet [2] reports a Dice of 0.75, which exceeds the vanilla nnUNet value of 0.68 reported in the paper but is beaten by the authors’ result of 0.78. Second, the results lack any measure of statistical variability, making it impossible to assess whether the observed differences are statistically significant.
On the minor side, the computational cost is not discussed despite the ablation study showing that the PET-only variant achieves 96% of the best Dice score with only about half the parameters (57%). Several methodological details are also missing: for MA-MIP, the number of projections and the angular step size are not specified; and for Dual-SE, neither the meaning of “dual” nor the role of “global” in the Global Average Pooling is clarified.
[1] Gatidis, Sergios, et al. “Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging. “ Nature Machine Intelligence 6.11 (2024): 1396-1405. [2] Andrearczyk, Vincent, et al. “Overview of the HECKTOR challenge at MICCAI 2022: automatic head and neck tumor segmentation and outcome prediction in PET/CT. “ 3D Head and Neck Tumor Segmentation in PET/CT Challenge. Cham: Springer Nature Switzerland, 2022.1-30.
- 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
Regarding the proposed architecture, it is not clear why the fused graph features are mapped back and residually added only to the PET bottleneck features, rather than to both PET and CT features. It would be worth clarifying whether this design choice is related to the observation that best performance was achieved with Q=PET in the cross-attention module; if so, it would also be informative to know whether, in the ablation study where Q=CT, the target of the residual addition was correspondingly changed to the CT bottleneck features.
As a minor editorial remark, the sentence “We adopt patient-level partitioning to prevent data leakage” appears to be repeated twice in the manuscript.
- 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 authors showcase a very effective and practical way to achieve good segmentation results by combining existing methodologies in a unified framework. However, the comparison against the state of the art did not seem fair in my opinion: the rebuttal will help to confirm whether my comment about published literature results is applicable or not.
- 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 rebuttal clarified several aspects such as computational cost, and statistical significance of results. The issue about the novelty/impact of the paper, which other reviewers identified as a weakness but I actually thought could represent a strength, has also been partially adressed. I am not convinced by the authors’ rebuttal for baseline results, although I better understand their point of view for reporting self-trained “vanilla” versions of the best-performing models (e.g. without post-processing) rather than pulling the best results from the literature. In my opinion, the rebuttal has sufficiently improved the quality to reach just above the bar for acceptance.
Review #2
- Please describe the contribution of the paper
The paper proposes CDFP-Net, a PET-CT tumor segmentation model that combines diffusion-pretrained feature priors with PET-guided asymmetric fusion of CT information. It also uses a 2.5D multi-angle projection scheme to reduce 3D computation while preserving global context. Together, these design choices improve segmentation accuracy and robustness across multiple datasets.
- 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 paper’s main strengths are that it presents a well-motivated and coherent method rather than a loosely assembled architecture, with novelty centered on PET-guided asymmetric graph fusion that uses metabolically salient PET regions to selectively pull complementary boundary information from CT. This is interesting because it reflects the distinct clinical roles of the two modalities better than standard symmetric fusion or simple concatenation. The use of a frozen diffusion-pretrained backbone with lightweight adapters is also a strong aspect, since it offers a practical way to exploit self-supervised priors under limited annotation and multi-center variation. In addition, the MA-MIP 2.5D representation is a smart efficiency choice that preserves global volumetric context without full 3D cost. These ideas are supported by broad evaluation across three public datasets, comparisons with strong baselines, and targeted ablation studies that validate the contribution of the pretraining, adapters, fusion design, and timestep choice.
- 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 main weaknesses of the paper are that several of its core ideas appear more integrative than fundamentally new: the diffusion backbone is based on standard DDPM pretraining, the asymmetric graph fusion is explicitly built on prior heterogeneous multimodal fusion work, and the MA-MIP strategy is close to prior projection-based PET segmentation methods already cited by the paper itself. In addition, while the application is clinically relevant, the evidence for clinical feasibility is still limited to retrospective benchmark results with standard segmentation metrics, without reader studies, workflow impact, or treatment-planning evaluation. The experimental section is solid but does not fully probe robustness to issues such as scanner variation, missing modalities, lesion type, or low-data regimes, and the efficiency claims are not directly quantified with runtime or memory comparisons. Finally, although the method performs strongly overall, it is not uniformly best on every metric, and the paper does not fully disentangle how much of the gain comes specifically from diffusion priors versus pretraining or architectural design more broadly.
- 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?
I gave the paper a 4 because it is a solid and credible contribution with clear technical merit, strong empirical results, and a well-motivated method for PET-CT tumor segmentation. The main positive factors were the strong overall performance across multiple datasets, the coherent integration of diffusion priors with PET-guided asymmetric fusion, and the supporting ablation studies. What kept it from a higher score is that several of the core ingredients are extensions or combinations of existing ideas rather than a fundamentally new paradigm, and the paper does not yet provide especially strong evidence of clinical deployment readiness or deeply probe robustness beyond the benchmark setting.
- Reviewer confidence
Very confident (4)
- [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’ reply to reviewer comments adequately addressed my concerns. I find the content to be worth presenting at MICCAI.
Review #3
- Please describe the contribution of the paper
The paper proposes CDFP-Net, a dual-stream PET-CT tumor segmentation framework that integrates three key components: 1.Diffusion-prior-guided representation learning via a frozen DDPM encoder with lightweight adapters, enabling transfer of self-supervised generative priors to segmentation under limited supervision. 2.Modality-asymmetric dynamic graph fusion, where PET features act as query anchors to selectively aggregate structural cues from CT using graph-based message passing and cross-attention. 3.MA-MIP (Multi-Angle Maximum Intensity Projection) strategy to convert 3D volumes into 2.5D representations, reducing computational cost while preserving global semantic context. The method is evaluated on three multi-center datasets (AutoPET, HECKTOR, PCLT20K) and shows consistent improvements over several SOTA baselines in DSC, HD95, and Recall.
- 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.Well-motivated integration of diffusion priors into segmentation: The use of a frozen DDPM encoder with lightweight adapters is technically sound and aligns with recent trends in leveraging generative priors. The use of v-prediction + SNR weighting is also a strong design choice that stabilizes representation learning. 2.Novel asymmetric cross-modal fusion formulation: The idea of PET as query and CT as key/value is conceptually meaningful and grounded in domain knowledge (functional vs anatomical signals). This is a clear improvement over naive concatenation, which is explicitly criticized in the paper. 3.Strong multi-dataset validation (important for MICCAI): Evaluation on three datasets with different characteristics is a major strength. This supports claims about robustness and generalization. 4.Clinically relevant task with clear translational potential: PET-CT tumor segmentation is highly relevant for radiotherapy planning (BTV delineation), which aligns well with MICCAI clinical impact expectations.
- 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.
Limited novelty in core methodological components (incremental contribution). The paper is largely a combination of existing ideas: 1.Diffusion priors for representation learning already explored (e.g., diffusion-based pretraining in medical imaging). 2.Graph-based fusion and asymmetric fusion already exist (e.g., AAHN, CMHGF). 3.MIP-based 2.5D modeling previously proposed. The novelty lies mainly in integration, not in fundamentally new methodology. Lack of comparison with strongest diffusion-based segmentation baselines. The paper compares with nnUNet, SwinUNETR, etc., but: No comparison with recent diffusion-based segmentation frameworks and, no comparison with foundation models (e.g., SAM-based medical variants, MedSAM, etc.). This weakens the claim of SOTA performance. Potential fairness issue in computational comparison. CDFP-Net uses 2.5D projections, while baselines are mostly full 3D models. This may introduce unfair efficiency vs accuracy trade-off. No FLOPs / runtime comparison is provided. Insufficient discussion of failure cases. The paper shows qualitative success (Fig. 2), but: No analysis of failure modes and No discussion of low PET signal edge cases beyond anecdotal examples. Weak clinical validation: No expert evaluation, No clinical endpoint validation, and No prospective or workflow-based evaluation. Claims of clinical applicability are somewhat overstated. Unclear details in diffusion pretraining setup: Size and diversity of pretraining dataset unclear, Whether PET and CT are trained jointly or separately not fully specified, and Impact of diffusion timestep selection (beyond one table) not deeply analyzed. Graph fusion complexity vs gain not justified: Added complexity (GCN + attention), and Gains are modest (~1–3% DSC in some cases). No cost-benefit analysis. Some references appear questionable or future-dated. Several citations (2025 arXiv, etc.) may indicate: Potential weak grounding in established literature, and Possible risk of citation inflation.
- 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
This is a solid and well-engineered paper with a meaningful application. However, to improve its competitiveness at MICCAI: 1.Clearly position the work against recent diffusion-based segmentation and foundation models. 2.Add runtime / memory / FLOPs comparison. 3.Include failure case analysis 4.Strengthen the clinical validation narrative 5.Clarify pretraining dataset and protocol
- 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 is strong in engineering, validation, and practical relevance but Weak in methodological novelty and positioning vs latest work. Key deciding factors: 1.Multi-dataset validation (important for MICCAI) 2.Well-designed ablations 3.Incremental novelty, Missing comparisons with strongest recent methods, Limited clinical validation. Overall, it is above threshold but not a clear 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 and meta-reviewer for their comments. We address the key concerns below. [Fair SOTA comparison, R1-Q1] The difference between our absolute scores and some published values mainly reflects dataset/protocol differences, not weak baselines. For HECKTOR, we used the expanded HECKTOR 2025 dataset (Ref. 14), which is larger and more heterogeneous than the 2022 version (from 854 to ~1,200 cases and from 9 to 11 centers); therefore, DSC values are not directly comparable. For AutoPET, standard DSC is sensitive to the 50.6% negative cases, and top entries often used dataset-specific post-processing, such as removing small connected components. To evaluate intrinsic representation ability, we did not use such post-hoc optimization and reported standard DSC, HD95, and Recall on positive cases. Under the official post-hoc oDSC protocol, CDFP-Net reaches 0.8488.[Variance and significance, R1-Q2] All metrics are averaged over five independent runs. Wilcoxon signed-rank tests show significant improvements across datasets (all p<0.05; e.g., p=0.017 on HECKTOR), supporting the robustness of the gains. [Computational cost, R1-Q3/R2-Q3/R3-Q3] CDFP-Net requires 133.14 GFLOPs, with 268.17 ms average single-sample inference time. Although PET-only is lighter, the full model improves HECKTOR DSC from 75.31% to 78.44% and reduces HD95 from 12.16 mm to 8.26 mm, showing a meaningful boundary/accuracy gain relative to the added cost. [Method details, R1-Q3/R2-Q3/R3-Q6] MA-MIP uses a 5° step, producing 72 projections per volume. To avoid confusion with the dual-stream backbone, “Dual-SE” is better described as two-stage SE, since it is deployed at the input and bottleneck stages. GAP recalibrates channels using global macro-context rather than local intensity spikes. PET and CT diffusion encoders are pre-trained separately on unlabeled modality-specific data. All datasets are split at the patient level into pre-training, fine-tuning, and test sets using a 7:2:1 ratio, with sliding-window cropping applied to all MA-MIP projections. [Contribution and novelty, R2-Q1/R3-Q1] We agree that CDFP-Net builds on diffusion pre-training, projection-based modeling, and graph fusion. Its contribution is a PET-CT-specific system aligned with three bottlenecks: limited annotation, high 3D cost, and PET-CT physical heterogeneity. The frozen diffusion backbone transfers self-supervised generative priors with lightweight adapters, while the modality-asymmetric dynamic graph mechanism uses metabolically salient PET regions as anchors to query CT boundary cues. This preserves modality heterogeneity and enables PET-guided, structure-assisted refinement beyond simple concatenation or symmetric fusion. [Clinical scope, R2-Q2/R3-Q5] We agree that the current evidence is retrospective and does not establish deployment readiness. Our claims should be interpreted as retrospective performance warranting further validation, not direct clinical feasibility. Future work will include prospective validation, reader studies, workflow assessment, and treatment-planning endpoint evaluation. [Baselines and positioning, R3-Q2] Our baselines include recent PET-CT fusion and medical SSL methods, including CIPA, AAHN, and MedCoSS, plus representative 3D segmentation models. We acknowledge that diffusion-based segmentation and foundation-model comparisons would further strengthen positioning and will explore them in future work. [References, R3-Q8] Available preprints have been updated to peer-reviewed versions where possible. Ref. 14 remains an arXiv citation because it is the required reference for the unpublished HECKTOR 2025 dataset. [Fusion logic, R1-Q4] We appreciate this important point. In the Q=CT ablation, the residual target was not changed to the CT bottleneck, which may understate the reverse-guidance setting. We will add supplemental experiments with CT-target residual addition to test the PET-guided, CT-assisted fusion hypothesis.
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 is a solid engineering paper with interesting integration of diffusion priors and cross-modal fusion. However, the contribution is largely incremental, the use of diffusion models is somewhat superficial, and several methodological and evaluation concerns limit its impact of this study.
all three reviewers found some concerns to be addressed or discussed, while two of them found merits too.
- 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.
it is a solid paper (engineering approach) with some incremental novelties. Since the weaknesses were addressed, the evaluations were improved, and application code of the paper is strong, and last but not least, two reviewers already locked their decisions to accept, the paper is a candidate for a poster presentation.
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.
A well presented paper with solid technical designs. The authors’ rebuttal addressed most reviewer concerns.
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.
This manuscript combines diffusion prior and cross-modal fusion into a unified framework for lesion segmentation in PET/CT images. The framework improves segmentation accuracy on multiple datasets, in comparison with reference baselines. The rebuttal addresses the reviewers’ concerns about statistical significance of experimental results, computational cost, technical contributions, and method details. Thus, the manuscript is recommended for acceptance.