Abstract

Pulmonary perfusion imaging is clinically important, but paired 3D CT-perfusion annotations are difficult to obtain. Semi-supervised learning (SSL) can exploit abundant unlabeled CT volumes, yet existing SSL methods are less effective for voxel-wise dense regression because continuous pseudo-labels lack reliable confidence estimates and strong consistency may amplify teacher noise. To address this, we propose ProtoReg, a semi-supervised framework for 3D non-contrast CT-to-perfusion regression with prototype-guided uncertainty modeling and sub-unit-aware learning. ProtoReg adopts a teacher–student framework with weak-to-strong consistency and maintains a learnable functional prototype bank to estimate uncertainty via prototype-induced entropy. Based on this uncertainty, we introduce a Dual-level Uncertainty-weighted Consistency Loss (DUW-CL), which re-weights consistency at both voxel and lung sub-unit levels to suppress noisy pseudo-label propagation. We further incorporate 1,000 external unlabeled CT volumes from public datasets to improve generalization without additional perfusion annotations. Experiments on two independent CT-perfusion cohorts show consistent improvements over strong SSL baselines, and ablations validate the effectiveness of prototype guidance, DUW-CL, and supervoxel-aware modeling. Code is publicly available at https://github.com/johnhamtom/ProtoReg_main.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0875_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/johnhamtom/ProtoReg_main

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhuYit_ProtoReg_MICCAI2026,
        author = { Zhuang, Yitao AND Pu, Yao AND Zhang, Zhenxi AND Yu, Peixin AND Li, Zihan AND Zhao, Mayang AND Ho, Wai Yin AND Chu, Jocelyn AND Chan, Gary Ka Wai AND Cai, Jing AND Ren, Ge},
        title = { { ProtoReg: Prototype Guidance and Dual-Level Uncertainty-Weighted Consistency for Semi-supervised Pulmonary Perfusion Regression } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes ProtoReg, a semi-supervised framework for 3D pulmonary perfusion regression from non-contrast CT. Its core contribution is adapting prototype-based learning (typically used for discrete segmentation) to a continuous dense regression task by quantizing perfusion values into functional bins. It further introduces a Dual-level Uncertainty-weighted Consistency Loss that leverages prototype-induced entropy at both voxel and lung sub-unit levels to mitigate noise in pseudo-labels.

  • 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.It effectively bridges the gap between discrete prototype learning and continuous voxel-wise regression.

    2.The evaluation includes very recent and competitive SSL baselines like UniMatch v2 (2025).

    3.The method is validated on both private and public datasets, showing feasibility for inferring lung function without radioactive tracers.

    4.Using lung sub-units (supervoxels) to provide regional consistency is a well-motivated approach for 3D medical volumes.

  • 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.While the authors compare against UniMatch V2, they omit discussion or comparison with prior prototype-based SSL methods like CPCL (Xu et al. , 2022) or UPCOL (Lu et al. , 2023). These are highly relevant as they also utilize prototype consistency in medical imaging.

    2.The framework appears to be a combination of prototype learning and the DyCON strategy. Since the uncertainty weighting (Eq. 8–10) follows DyCON closely, the authors should clarify their specific technical innovations beyond this combination. A direct experimental comparison with DyCON is also missing.

    3.In Equation 8, the parameter β is used without a clear definition or initial explanation in the preceding text.

    [1] Xu Z, Wang Y, Lu D, et al. All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation[J]. IEEE Journal of Biomedical and Health Informatics, 2022, 26(7): 3174-3184. [2] Lu W, Lei J, Qiu P, et al. Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation[C]//International conference on medical image computing and computer-assisted intervention. Cham: Springer Nature Switzerland, 2023: 662-672.

  • 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.

    (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 addresses a challenging and clinically valuable task. I appreciate the inclusion of 2025 baselines. However, the lack of comparison with more specialized prototype-based SSL methods and the strong reliance on the DyCON formulation without a direct comparison are concerning. If the authors can clarify their specific technical contributions over DyCON and discuss/compare against the aforementioned prototype methods in the rebuttal, I would be open to increasing my score.

  • 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 have successfully addressed my concerns in the rebuttal. The authors provided a clear theoretical distinction from DyCON regarding continuous 3D regression. They also properly clarified the notation of β and the omission of CPCL/UPCOL due to task differences. I maintain my positive recommendation to Accept.



Review #2

  • Please describe the contribution of the paper

    1.This paper proposes a semi-supervised framework for CT-to-perfusion synthesis, named ProtoReg, that addresses the paired CT–perfusion acquisition problem. 2.This paper tackles the continuous pseudo-label prediction problem by using prototype-induced entropy. 3.The results show consistent improvements across all 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.

    1.This paper introduces a framework for a semi-supervised CT-to-perfusion task. This topic is novel and clinically meaningful. 2.The incorporation of dual-level (voxel-level and sub-unit-level) is reasonable and validated to improve the overall performance. 3.The experiment results demonstrate robust performance across both cohort datasets, and the evaluation metrics employed are appropriate for assessing perfusion data.

  • 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 notations in the method are sometimes difficult to follow (e.g., v first appears in sec. 2.2, but is first defined in sec. 2.3). The manuscript would benefit from clarity in notation. 2.The authors state that the model initially focuses on low-uncertainty regions and later shifts to refinement with milder entropy-guided weighting. However, in Eqs. (8) and (9), the notation w continues to assign higher weights to low-entropy predictions throughout training. Please clarify what is meant by “shifts to refinement with milder entropy-guided weighting” and justify the use of the geometric decay schedule. 3.The proposed DUW-CL appears closely related to UnCL in DyCON, with the main difference being a dual-level formulation. The authors should more clearly articulate the novelty and distinction of their approach. 4.Prototype-induced entropy for reliability assessment may not be entirely novel, as similar ideas have been explored in prior works such as [1,2]. Nevertheless, this might be the first work to introduce such a reliability assessment strategy in the CT-to-Perfusion synthesis task. [1] A unified uncertainty network for tumor segmentation using uncertainty cross entropy loss and prototype similarity (Diao et al., 2022) [2] Comprehensive Evaluation of Prototype Neural Networks (Schlinge et al., 2025) 5.The original volumetric sizes of the datasets are not specified. Is it appropriate to uniformly resize all data to 128128128? Please provide details on the preprocessing pipeline and discuss whether interpolation affects the integrity of CT and perfusion data. 6.The authors claim consistent improvements on both private and public CT-perfusion datasets. However, the reported results only include private datasets (Cohorts A and B). Please clarify whether experiments were conducted on public datasets or only for training.

  • 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 provide sufficient information for 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?

    This paper addresses a clinically meaningful challenge in CT-to-perfusion synthesis. However, several issues remain unresolved in the manuscript. In particular, the overall method design appears to be largely adapted from prior work in other domains, and the originality and domain-specific contributions of this study should be more clearly articulated.

  • 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.

    Lack of ablation studies, but agree with the clinical practice and the statistical analysis results.



Review #3

  • Please describe the contribution of the paper

    This paper proposes ProtoReg, a semi-supervised framework for 3D CT-to-perfusion regression (predicting SPECT-like perfusion maps from non-contrast CT). The framework uses a teacher-student setup with weak-to-strong consistency on a 3D U-Net backbone. The main contributions are: (1) a prototype bank where K functional prototypes are learned from labeled data by quantizing perfusion values into bins, (2) prototype-induced entropy as a confidence signal for continuous pseudo-labels (since regression has no natural confidence score), and (3) DUW-CL, a Dual-level Uncertainty-weighted Consistency Loss that re-weights voxel-wise consistency using both voxel-level and lung sub-unit (supervoxel) uncertainty. They also incorporate 1,000 external unlabeled CT volumes from public datasets. Experiments on a private cohort (233 patients) and a public cohort (100 studies) show improvements over MT-Reg, FixMatch-Reg, UniMatch-Reg, and UniMatchV2-Reg.

  • 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.Underexplored problem: Semi-supervised learning for dense regression in 3D medical volumes is a real gap in the literature, most SSL work focuses on classification or segmentation. CT-to-perfusion synthesis is also clinically meaningful since SPECT acquisition is expensive. The application is well-motivated. 2.Solid baseline selection: The paper compares against MT-Reg, FixMatch-Reg, UniMatch-Reg, and UniMatchV2-Reg, all adapted to dense regression with the same backbone and protocol. Including UniMatch V2 as a baseline is the right call and reflects awareness of the latest work. 3.Ablations included: Table 3 covers prototype bin count (K=2,5,10), comparison against fixed-threshold filtering (th_consist=0.90/0.95), and the contribution of LSV-level weighting. This breaks down the design choices reasonably well. 4.External unlabeled data scaling: Adding 1,000 external CT volumes from public datasets and showing it still works is a practical demonstration. The 100% labeled + 1000 external CT setting is realistic for clinical scenarios where unlabeled CTs are abundant.

  • 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.Improvements are very small relative to standard deviations. On Cohort A 100%, ProtoReg gets ρ=0.800±0.090 vs UniMatchV2-Reg at 0.797±0.100, a 0.003 difference with std devs of ≈ 0.1.On Cohort A 20%, ProtoReg is 0.778±0.110 vs UniMatchV2 at 0.776±0.110, essentially identical. The margins are well within run-to-run variance. No paired statistical tests are reported, so it’s impossible to know if these differences are meaningful. 2.Semi-supervised methods sometimes hurt performance. On Cohort B with 100% labels + 1000 unlabeled CT, MT-Reg (ρ=0.802), FixMatch-Reg (0.797), and UniMatch-Reg (0.801) all perform worse than the fully supervised baseline (ρ=0.810). ProtoReg only barely beats Sup-only (0.814 vs 0.810). The paper doesn’t address this finding, if external unlabeled data hurts most SSL methods on this cohort, that’s a significant negative result that needs to be addressed. 3.Method complexity vs gains. The full pipeline includes: teacher-student EMA + prototype bank + prototype classification loss + prototype-induced entropy + voxel-level uncertainty + LSV (supervoxel) uncertainty + dual-level consistency weighting + entropy regularization + geometric decay schedule. That’s a lot of moving parts for ~0.3% Spearman improvement over UniMatchV2.The ablation shows the LSV-level weighting helps (0.778 vs 0.774), but again the margin is small relative to noise. 4.Limited ablation scope. Ablations are only on Cohort A under the 20% labeled setting. The 100% labeled setting and Cohort B are not ablated. So we don’t know if the same components matter when labels are abundant or on a different cohort. This makes it harder to argue the design generalizes. 5.Some experimental details are unclear. It’s not clear if results are from single runs or averaged across multiple seeds. The train/val/test split protocol isn’t fully specified. The number of test cases per cohort isn’t stated explicitly.

  • 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

    1.Are the reported numbers from single runs or averaged over multiple seeds? Can you provide paired statistical tests between ProtoReg and UniMatchV2-Reg? 2.On Cohort B with 100% labels, why do most SSL methods underperform the supervised baseline? Does this suggest the unlabeled data is hurting more than helping in some settings? 3.Can you provide ablations on Cohort B and on the 100% labeled setting to confirm the design choices generalize?

  • 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 tackles a meaningful underexplored problem (semi-supervised dense regression for CT-to-perfusion) and the baseline selection is appropriate, including the latest UniMatch V2.However, the improvements over the strongest baseline are very small relative to standard deviations, often a fraction of a percent on Spearman ρ with std devs around 0.1 and no statistical significance tests are reported. The Cohort B results show that most SSL methods actually hurt performance compared to the supervised baseline at full label budget, which is a notable finding the paper doesn’t engage with. The method has many components but the ablation is limited to one cohort and one labeling ratio, so the generalization of the design choices is unclear. The core contributions are reasonable but heavily inspired by DyCON. On balance, the application is interesting and the experimental setup is thoughtful, but the small margins, lack of statistical tests, and unaddressed negative result tip this toward rejection. Statistical significance testing, broader ablations, and a discussion of the Cohort B finding would substantially strengthen the paper.

  • 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.

    My pre-rebuttal recommendation was Weak Reject. The rebuttal engaged with my concerns thoughtfully and clarified the DyCON distinction, but the substantive issues remain: (a) statistical significance on the primary metric is not established even with the new tests, (b) the core mismatch between method complexity and effect size persists. The application is interesting and the work is competent, but the contribution as evidenced doesn’t clear the MICCAI bar. I maintain my recommendation to reject.



Author Feedback

We thank the reviewers and AC for their comments. We will release the code and provide the public dataset (Cohort B) download link upon acceptance. Below are our responses. A. Method novelty/description. [R2-W1, R3-W4] Related methods. Resp: We thank the reviewers for pointing out CPCL, UPCoL, DyCON, and related uncertainty-estimation methods. These studies mainly rely on discrete category-level signals, e.g., class predictions or probability maps. Our task is continuous voxel-wise regression, where pseudo-targets have no categorical confidence scores. Their uncertainty formulations are therefore not directly applicable as baselines. We will cite these works and clarify differences from ProtoReg. [R2-W2, R3-W3] Difference from DyCON. Resp: ProtoReg differs from DyCON in two key aspects. First, DyCON estimates uncertainty from discrete segmentation probabilities, while ProtoReg models uncertainty from continuous functional values through functional prototypes. This bridges discrete prototype learning and continuous voxel-wise regression, providing reliability signals for pseudo-targets without class probabilities. Second, DyCON performs voxel/pixel-level uncertainty weighting, while DUW-CL computes reliability-weighted consistency at both voxel and anatomically meaningful lung-sub-unit levels. This captures regional functional continuity and suppresses fragmented pseudo-label noise. We will add these clarifications. [R2-W3, R3-W1, W2] Notation and weighting schedule. Resp: We will check all notations and further define β before Eq. (8), clarifying its role in DUW-CL. In DUW-CL, w always gives higher weights to low-entropy predictions, while β controls entropy-weighting strength. As β decays, the gap between low- and high-entropy weights becomes smaller, smoothly expanding training from highly reliable regions to broader consistency refinement. B. Data/preprocessing clarification. [R3-W5, R3-W6, R5-W5] Preprocessing and dataset split. Resp: We will further add preprocessing and split details. Paired volumes were pre-registered and processed using a standardized pipeline [16,17]. Original CT/SPECT sizes are approximately 512×512×200 and 128^3.Only the pre-segmented lung ROI was resampled to 128^3, preserving CT–SPECT spatial correspondence and minimizing task-relevant information loss. Cohorts A/B use separate fixed 80/20 train/test splits; 25% of training is used for validation, and labeled/unlabeled splits are applied only to the remaining training data. Test sets contain 48/20 cases, and results are per-case mean±SD. C. Results/statistics clarification. [R5-W1, A1] Statistical Significance. Resp: We have conducted metric-wise paired statistical tests against UniMatchV2-Reg and will report p-values in the revision. ProtoReg achieves statistically significant improvements (p<0.05) over UniMatchV2-Reg on most metrics under most settings. Under Cohort A 20%, ProtoReg reaches p<0.05 on SSIM, MAE, and FL20.On Cohort B, most metrics also reach p<0.05 under both 100% and 20% settings. [R5-A2, W2] SSL performance on Cohort B. Resp: We will clarify this in the revision. Several SSL baselines underperform Sup-only in Cohort B 100%, likely due to domain shift or pseudo-label noise when full-label supervision is strong. ProtoReg mitigates this through prototype-induced uncertainty and dual-level weighting. [R5-W3, W4, A3] Ablation. Resp: We will clarify the ablation findings. LSV-level weighting achieves statistically significant improvements in MAE, SSIM, and FL20.These metrics better reflect functional-region consistency and clinically relevant spatial fidelity, while Spearman ρ mainly measures global rank correlation and is less sensitive to local structural improvements. Due to page limits, we reported Cohort A 20% ablations, which best reflect the low-label reliability problem. We also conducted 100%-label ablations with consistent component benefits, and will include them with broader Cohort B analyses in an extended version.




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.

    All reviewers are inclined to finalize their decisions following the rebuttal. The authors are advised to address the concerns raised by Reviewers 3 and 4.

  • 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.

    Two reviewers found the rebuttal fully addressed their concerns and recommend acceptance. One reviewer remains concerned that the gains over strong baselines are modest and may not justify the added complexity. While these concerns are valid, the work addresses an important problem, includes competitive comparisons against recent methods, and the rebuttal clarified the methodological contribution and provided additional statistical analysis. On balance, the positive assessments outweigh the remaining concerns.



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.

    After the rebuttal, two of the reviewers are satisfied with the authors’ answers and recommend acceptance. The third reviewer remains concerned about the low improvement achieved by the method for its added complexity. After reviewing the paper and rebuttal, the AC feels that the positive points of the paper outweigh the negative ones, and recommends acceptance. However, the authors are strongly encouraged to update their paper following reviewers’ comments and suggestions, in particular, regarding the statistical significance of results.



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.

    After the rebuttal, one reviewer raised the score to acceptance, and most reviewers are positive. The main contribution lies in the clinical importance of the semi-supervised regression task and the sound solution proposed for linking discrete prototypes with continuous voxels. The authors also clarified the differences from the DyCON model. Overall, this work is potentially beneficial to the relevant community. The authors are encouraged to improve the manuscript in the camera-ready version further, and acceptance is recommended.



back to top