List of Papers Browse by Subject Areas Author List
Abstract
4D flow MRI provides rich hemodynamic information by encoding blood-flow velocities in the image phase; however, image quality highly depends on the velocity-encoding (venc) setting—high venc reduces the velocity-to-noise ratio (VNR), whereas low venc improves VNR but increases the risk of phase wrapping, potentially introducing substantial bias in downstream flow quantification. We propose PUDIP-Flow, an unsupervised and segmentation-free phase-unwrapping method for aortic 4D flow MRI based on the deep image prior (DIP). The method formulates phase unwrapping as an optimization problem that enforces consistency with wrapped spatiotemporal phase gradients, and solves it by fitting a 3D U-Net to a single scan without paired training data. To reduce reliance on manual segmentation of regions of interest, we further introduce an automated weighting mask derived from phase-contrast MR angiography features and temporal variance, thereby suppressing contributions from stationary background tissue. Experiments on retrospective and prospective aortic datasets show that PUDIP-Flow improves unwrapping accuracy and robustness compared with established baselines, with particularly strong performance when manual segmentations are unavailable.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1536_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AssociatedPrimeIdeal/PUDIP-Flow
Link to the Dataset(s)
N/A
BibTex
@InProceedings{RenYuy_Unsupervised_MICCAI2026,
author = { Ren, Yuyang AND Zhou, Zijian AND Hu, Peng},
title = { { Unsupervised and Segmentation-Free Phase Unwrapping for 4D Flow MRI Using Deep Image Prior } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This manuscript proposes PUDIP-Flow, an unsupervised phase-unwrapping method for aortic 4D flow MRI using the Deep Image Prior (DIP). The authors formulate unwrapping as a gradient-domain optimization problem, fitting a 3D U-Net to a single scan to enforce spatiotemporal continuity. Comparative studies were conducted to compare PUDIP-Flow with LAP, NRPS, and PUMA algorithms, and the results demonstrated that the proposed method led to desirable results.
- 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.
Overall, the logic flow of this manuscript is clear, and it is well-organized and easy to follow. The results is also desirable compared the other traditional comparative algorithms.
- 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 proposed PUDIP-flow is merely an application of the well-established DIP algorithms to the MRI phase unwrapping problem, and the innovation in the methodological point of view is low for MICCAI, considering that DIP has already been applied to MRI phase unwrapping, e. g. , DIP-UP method https://www. mdpi. com/2078-2489/16/7/592.
- The authors showed a limited literature review in the manuscript about the more recently proposed phase unwrapping algorithms, especially the deep learning ones, like PHU-Net, Phase-Net, PhaseNet2.0, PU-GAN, et al.
- Although the proposed PUDIP-Flow is oriented to the 4D flow data; however, there are no special or innovative designs for catering to the 4D spatial-temporal nature of the data. Instead, the authors only used a multi-channel U-net, which is basically still a 3D algorithm.
- There are also some efficiency concerns for the proposed method, which is nearly 8 minutes per scan, which is much longer than traditional Laplacian or path-based algorithms, and should also be much longer than the above-mentioned deep learning-based algorithms, especially the supervised learning ones.
- The loss function (eq. 6) design is also too simple. As the authors described in the manuscript, Itoh’s condition is only valid when the gradient of flow is limited; however, they did not show this is the case for their applications. In addition, this theory is pretty out-of-date (proposed in 1982); indeed, there are many state-of-the-art equations building relationships between the wrapped phase and the potential unwrapped solutions (some of them are also the basis for the Laplacian unwrapping).
- 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.
(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?
My major concerns are about the technical novelty of the manuscript. DIP is a well-established technique in CV and has been successfully applied to many areas in the medical imaging field already. I do not think the “A+B” type of innovations is good enough for MICCAI. Specifically, the DIP-UP / PHU-Net / PhaseNet papers I recommended to the authors are just an arbitrary search on Google Scholar using the keywords “DIP” + “Phase Unwrapping” + “MRI”, while the authors did not present any introductions, comparisons, or discussions. More detailed commnets can be found in the above “drawback” part.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The DIP-UP method I mentioned in my previous comment is just a quick search on Google Scholar using “DIP + phase unwrapping”, and there is also some other good fully unsupervised phase unwrapping method: “Robust Phase Unwrapping via Deep Image Prior for Quantitative Phase Imaging” (QPI) published at IEEE Transactions on Image Processing 2021 (DOI: 10.1109/TIP.2021.3099956), which is nearly the same topic as this manuscript from the methodological point of view. So, this work is definitely not the first fully unsupervised, segmentation-free DIP pipeline for phase unwrapping.
In addition, it is really easy to transfer DIP-UP / QPI to 4D flow data, since the major methodological improvement is a 4D loss, while the model for iterative tuning is still a 3D conventional U-net, as used in DIP-UP and QPI.
So, the novelty of this manuscript is just marginal and limited, and their literature review work is not enough as well, without comparing to any deep learning-based comparative methods, considering it is very easy to transfer these methods for 4D flow data (e.g., conduct phase unwrapping for each time point one by one).
Review #2
- Please describe the contribution of the paper
The paper introduces a phase unwrapping approach for 4D flow MRI data based on the Unet and a loss function which formulates Itohs condition combined with a weighting which considers the temporal variation of flow velocities. The authors evaluate their approach against other state-of-the-art methods with regard to noise and the amount of voxels with phase wrap.
- 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 authors present a deep learning based approach which implements concepts from earlier segmentation and optimization-based phase unwrapping approaches. The concepts are well motivated clearly described and compared to the state of the art.
- 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 authors do not present an ablation study to assess the benefit of the suggested weighting concept.
- 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.
(5) Accept — should be accepted, independent of 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 for a clinically relevant problem. The implementation is described in such a way that it can be reproduced and compared against the state of the art.
- Reviewer confidence
Very confident (4)
- [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 #3
- Please describe the contribution of the paper
This paper proposes an unsupervised and segmentation-free phase-unwrapping method for aortic 4D flow MRI based on the deep image prior. This method adopts a 3D U-Net to generate unwrapped phases corresponding to the 4D flow MRI via an unsupervised training manner. Retrospective and prospective experiments demonstrate the effectiveness of the proposed method.
- 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 major strength of this paper is the idea of using the unsupervised and automatic segmentation mask to generate the unwrapped phases of 4D flow MRI.
- 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 rationality and novelty of using a 3D U-Net to generate the unwrapped phases should be further clarified. 2.The compared methods, i.e., Refs. [1], [5], and [9] were proposed more than 10 years ago. As a result, the current experimental results are partially convincing. The comparison should be performed with more up-to-date methods. 3.The unsupervised training, validation, and test details of the proposed method are missing. These details are important for further understanding of this method.
- 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?
Please refer to the listed major weaknesses of this paper.
- 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.
N/A
Author Feedback
We thank the reviewers and AC. We address the major concerns below. Q1.Novelty vs. DIP-UP and recent DL phase-unwrapping methods (R1, R3, Meta). 4D flow phase unwrapping differs structurally from prior settings: the target is a phase-difference Φ = ∠(I_FE · I^*_FC) (Eq. 1) — stationary tissue gives Φ ≈ 0, air is pure noise, and signal concentrates in vessels — unlike static GRE field-map unwrapping where every brain voxel encodes B0.The problem also spans three velocity-encoding directions and cardiac frames.
Despite the name, DIP-UP [Zhu 2025] REQUIRES supervised pre-training (PHU-Net3D/PhaseNet3D on ~28,800 paired patches with synthetic GT) before DIP refinement, uses a Laplacian loss, and targets static 3D brain GRE for QSM with no temporal or velocity-encoding component. PUDIP-Flow is unsupervised, uses a 4D spatiotemporal gradient-domain L1 loss, and targets the 4D flow phase difference. PHU-Net/PhaseNet 2.0/PU-GAN are likewise supervised, for 2D optical phase or static MRI.
Our contributions therefore stand: (i) the first fully unsupervised, segmentation-free DIP pipeline for 4D flow phase unwrapping with a 4D spatiotemporal gradient objective (Eq. 6); (ii) an automated PCMRA × temporal-std weighting mask (Eq. 8) exploiting the phase-difference structure; (iii) systematic validation across vencs and SNRs.
Q2.Baselines and uncompared recent methods (R1, R3, Meta). LAP [9], NPRS [1], PUMA [5] are the standard 4D-flow baselines, forming the public 4DFlow-unwrap benchmark [6] whose reference implementations we use. Two 4D-flow-specific recent methods are not included: (a) Berhane et al. [4] is supervised, requiring large paired wrapped/unwrapped data — the gap our method addresses; (b) CWLS [Ref. 19] is neither open-sourced nor reproducible from the paper. The 2D networks R1 lists target a different problem and need paired 4D flow data to retrain.
Q3.”Just a 3D U-Net” — the 4D design (R1, R3). Stacking temporal frames as output channels of a 3D U-Net is deliberate: the architecture stays lightweight while temporal continuity is imposed by the 4D finite-difference operator ∇ : R^N → R^4N in the loss (Sec. 2.2). Table 2 confirms this is load-bearing — replacing spatiotemporal TV with spatial-only TV raises nRMSE. Temporal information also drives the mask via std_t(P) (Eq. 8), most robust in Table 3. Q4.Ablation of the weighting concept (R2). Tables 2 and 3 include the relevant ablations: width Nc, L1 vs. L2, spatial-only vs. spatiotemporal TV, and four weighting masks (Pstd, P, mean-P, all-ones). Pstd is the most robust across vencs.
Q5.Computation time (R1). The ~8 min figure is on a single GPU without ROI cropping; both factors dominate. The three velocity-encoding directions are independent and can parallelize across GPUs. This post-processing is offline; a few minutes per scan fits clinical workflow and does not affect bedside use.
Q6.Reproducibility (R3, Meta). Sec. 2.3 lists the experimental settings; baselines use the public 4DFlow-unwrap repo [6]. Code will be released upon acceptance.
Q7.Itoh’s condition (R1). We acknowledge |∇Ψ| ≤ π is not formally proven, and under very severe wrapping it can fail locally and degrade accuracy — visible as the rising nRMSE at the lowest vencs in Tables 1–2.For our experimental range (30–90 cm/s), however, the condition holds in the vascular ROI on physical grounds: aortic flow follows viscous incompressible Navier–Stokes, yielding fields smooth in space and time at our scales, so discrete gradients of Ψ stay below π. On “outdated” alternatives: Eq. (4) underpins LAP [9, 2016], Poisson [14, 1995], WLS [19, 2021] — methods spanning three decades, including R1’s own suggested baselines; the Schofield–Zhu Laplacian identity R1 alludes to (basis of LAP, already our baseline) is a reformulation of the same local-smoothness premise, more noise-sensitive (second-order). Local violations are absorbed by the L1 loss (Table 2), DIP prior, std_t mask, and congruence step (Eq. 9)
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 manuscript presents a clinically relevant unsupervised phase-unwrapping method for 4D flow MRI, with promising results especially in the segmentation-free setting. However, the novelty over existing DIP-based and deep-learning phase-unwrapping methods is not clearly established, and the comparison to recent MRI phase-unwrapping methods appears incomplete. The rebuttal should focus on clarifying the specific methodological novelty, justifying the choice of baselines, and providing missing implementation details needed for reproducibility.
- 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 reviewers found the problem clinically relevant and the results promising, especially for phase unwrapping without manual segmentation. The rebuttal clarified the main concerns about novelty, baseline choices, the weighting-mask ablation, and implementation details.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper received after rebuttal two “accept” and one “reject” recommendation from the reviewers. One reviewer keeps having significant reservations regarding novelty and performed literature review that could not be sufficiently resolved in the rebuttal. While the topic is certainly important and - as pointed out by one of the reviewers - has relevance for other areas relying on phase processing, search for similar literature indeed reveals some closely connected studies that make the current paper appear incremental in nature.
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.
Reviewers raised concerns on its novelty since simiilar ideas has been published somewhere else. However, it is still worthy of exploring its usage in a new clinical scenario
