List of Papers Browse by Subject Areas Author List
Abstract
High-resolution (HR) magnetic resonance imaging (MRI) offers exceptional visualization of human tissue but is often limited by hardware constraints. While recent super-resolution (SR) methods leveraging learned codebooks have shown promise, they often overlook the rich anatomical priors inherent in MRI data.
To address this, we propose a probabilistic prior-guided anatomical alignment for MRI super-resolution (PGASR) method that incorporates anatomical knowledge into the SR process. Specifically, we first introduce an anatomical-conditioned codebook generation (ACG) module that generates rough anatomical structure maps by extracting the regions of interest from MRI slices. These maps are used as anatomical conditions for the discrete codebook generation. Then, to better exploit information between MRI slices, we propose a prior matching alignment (PMA) module that aligns the codebook index matching probabilities between adjacent slices, as well as across low-resolution (LR) and high-resolution (HR) domains, thereby reducing the loss of image details. We validate the effectiveness of the proposed PGASR method with the public MRI dataset IXI. The experimental results demonstrate that PGASR outperforms state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5109_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LuoYiw_Probabilistic_MICCAI2025,
author = { Luo, Yiwen and Tang, Xiaoying and Yuan, Yixuan},
title = { { Probabilistic Prior-Guided Anatomical Alignment for MRI Super-Resolution } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces an ACG module—leveraging anatomical priors extracted by the SAM to condition discrete codebook construction and thereby better preserve tissue boundaries in super‑resolution reconstruction. It also presents a PMA module that enforces KL divergence alignment of pre-trained codebook matching probabilities across adjacent slices and between LR and HR domains to strengthen volumetric consistency.
- 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.
This work proposes a deep representation approach for modeling inter-slice relationships, guided by structural priors extracted using the SAM model.
- 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 claim that their proposed method performs well in detail recovery. However, as shown in Figure 4, the improvements of the proposed method over other methods are limited. The authors lack further explanation for this conclusion. Additionally, the quantitative metrics for the IXI dataset presented in Table 1 and Table 2 do not include statistical standard deviation metrics, making it impossible to evaluate the stability and statistical significance of the proposed method.
- Please rate the clarity and organization of this paper
Poor
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
- The authors should provide additional experiments to thoroughly demonstrate the feasibility and superiority of the proposed method. Consequently, relying solely on the results presented in Figure 4 is highly limited. Furthermore, the authors should supplement the quantitative metrics with statistical measures, such as standard deviation, to adequately showcase the robustness of the proposed method.
- Clinical data should be further supplemented to thoroughly demonstrate the clinical applicability of the proposed method. Moreover, the method presented in this study incorporates an excessive number of loss functions to optimize model parameters, and the stability of this approach requires further validation. The authors should explore whether the model can be streamlined while preserving its core architecture, as this is typically beneficial for enhancing model stability.
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The results presented in the work demonstrate limited improvement compared to other competing methods.
- 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 addressed some of my concerns.
Review #2
- Please describe the contribution of the paper
The authors propose a prior informed super-resolution (SR) approach for brain MRI. The model is trained on 2D low resolution slices with corresponding high-resolution ground truth. Prior anatomical information is obtained via tissue segmentations of the slices using the foundation segmentation model SAM (segment-anything model). This prior is encoded and quantized in an “anatomical” codebook. For super-resolution of LR slices, the slices are encoded and quantized. Codebook consistency between LR and HR domain as well as neighboring LR slices is enforced. The final HR slice is then obtained via the decoder G.
- 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: Novel approach to integrate anatomical priors using a general segmentation model (SAM). 2: Well-founded assumptions on codebook regularization enforcing closeness between neighboring slices and consistency between LR and HR domain. 3: Strong results with comparison to numerous baselines. 4: Detailed architecture, training- and hyperparameters. 5: Detailed ablation study.
- 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: Methodology is not always clear. Questions include: 1.1: Assuming the model operates on HR and LR slices, I_HR is denoted as R^(HxWx3). It is unclear what the last dimension represents given MRI are gray-scale images. 1.2: It is unclear how L_adv works and on which input the loss function operates on. Also, f the authors employ a discriminator, it is missing in Figure 2. 1.3: It is unclear if Encoder Ê and Feature Extractor E are “identical” or not (see Page 5 in the manuscript: “…feature extractor E (identical to the encoder Ê)”) Please elaborate more on the relation of the two encoders given they operate on HR and LR respectively. 1.4: The authors are encouraged to better clarify Fig 2. Following the pipeline is difficult. E.g, I_HR is supposed to represent the composition of Î_HR and S_HR but shows suddenly completely different images. 1.5: In Figure 2, top, Anatomical Prior Extraction: It is unclear to the reviewer what the bottom image represents (tissue boundaries?), how it is obtained, and why this step is necessary. 1.6: It is hard to disentangle which parts are employed and/or trained for Stage1, Stage2, and Inference. 2: Evaluation: 2.1: While the evaluation is thorough and convincing explicit evaluation of the proposed codebook regularization are missing. Similar like in Fig. 1, it would be interesting to see how the proposed changes affect the codebook activations.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Well structured paper with a novel codebook approach for super-resolution of MRI. Room for improvements includes mainly clarity of the architecture and methodology.
- 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.
The authors adequately addressed most of the concerns and clarified several methodological aspects of the paper. With these changes integrated to the manuscript, the work will be valuable for the conference.
Review #3
- Please describe the contribution of the paper
This paper proposes a novel codebook-based super-resolution approach for MR images.
- 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) Anatomical priors, obtained by SAM, are incorporated into images to enhance the input I_HR. (2) The paper proposes two alignment loss terms, i.e., L_align_s and L_align_d, to ensure the consistency between adjacent slices and reduce the domain gaps between HR and LR, which sounds novel to me. The ablation study also proves the feasibility of these designs. (3) The writing and organization of this paper are good.
- 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) MINet [3] and T2Net [2] are multi-contrast MRI SR methods, i.e., they require an auxiliary modality (“a reference image”) to provide guidance and assist in the super-resolution of target images. How are they utilized in the comparative experiments without references? (2) MR images are gray-scale and thus have only one channel. However, the authors state that “for a given HR image \hat{I}_HR \in R^HW3…” (in Page 4). Is the \hat{I}_HR equal to “Concat(I^{t-1}_HR, I^t_HR, I^{t+1}_HR)”? Please clarify. (3) Since the segmentation labels should be one-hot and thus discrete, how are the segments {s_i} aggregated into S_HR? What is the numerical range of S_HR, e.g., [0, 1]? (4) Minor issue: Should the “L_slign_d and L_slign_s” on page 3 be “L_align_d and L_align_s”?
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Despite some drawbacks, I believe the model’s novelty and the overall good writing of the paper could be reasons for acceptance.
- 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.
The rebuttal addresses my most concerns.
Author Feedback
We thank all reviewers for their invaluable comments and approval that our method is novel and effective. Our code will be released to ensure reproducibility. Common questions are first answered, followed by responses to individual comments.
R1&R2&R3: Evaluation of Codebook matching A: We further validate our codebook matching process by calculating KL divergence between HR and LR codebook activations, which quantitatively indicates the effectiveness of the alignment loss. For ×2 and ×4 LR, the KL divergence drops from 0.0597 and 0.0861 (without alignment loss) to 0.0413 and 0.0530 (with alignment loss), demonstrating that the alignment loss significantly improves codebook matching across domains.
R1&R2: Network components and losses for different stages A: Stage1 creates the conditioned codebook by jointly optimizing the encoder Ê, codebook Z, decoder G, and discriminator D through adversarial learning with L_stage1. In stage2, only the feature extractor E and spatial cross-attention module are optimized using L_stage2, while Z and G are frozen and employed for super-resolution. In inference, the same modules as in stage2 are employed, with all parameters fixed.
R1 suggests streamlining model for better stability. Stage 1 uses standard adversarial losses, while stage 2 employs two custom alignment losses to supervise codebook matching and uses L1 for reconstruction. While reducing the number of losses may improve stability, each loss in our framework serves a distinct and necessary purpose, so further reduction may not be directly applicable to our approach.
R2&R3: Input of HR size A: The original MR images are grayscale (H×W). However, we found that directly using them for codebook generation led to unstable training. Thus, we duplicated the channel to convert images to RGB (H×W×3) to address this issue.
R1: Robustness and effectiveness of model A: All results in table 1 are averaged over three runs. We report the standard deviations as follows. For ×2, PSNR, SSIM, and LPIPS are 36.7642 ± 0.2536, 0.9706 ± 0.1047, and 0.0487 ± 0.0087; for ×4, they are 30.6844 ± 0.3077, 0.8972 ± 0.1839, and 0.0924 ± 0.0095. These results confirm the stability and consistency of our method. The generated priors in Fig.2 clearly delineate anatomical regions, and Fig.4 shows that our method achieves the highest PSNR. Due to constraints in rebuttal, we will add experiments on additional clinical data in final version.
R2. How L_adv works A: The adversarial loss L_adv is used in conditioned codebook generation (VQGAN-like architecture), where the discriminator distinguishes real from generated HR images. For simplicity, the discriminator is initially omitted from Fig. 2. we will include it for clarity.
R2: Clarification on encoder Ê and E A: Encoder Ê and E share the same architecture but are trained independently in stage1 and stage2. Both process images of the same size, as LR images in stage2 are degraded and resized to HR dimensions. Using identical architecture ensures codebook matching between HR and LR representations.
R2: Clarification on I_HR, Î_HR and S_HR in fig. 2 A: In Figure 2, we use edge images to highlight the generated structural information, but this may have caused confusion. I_HR is derived by the concatenation of Î_HR and S_HR. This will be updated.
R3: Implementation Details of MINet and T2Net A: We followed the original protocols of each method. IXI provides multiple modalities for each subject. For MINet, we used T1-weighted HR to guide T2 super-resolution. For T2Net, we used noise-free images as X’_LR, which acts as a reference that provides rich anatomical structure.
R3: Obtain of S_HR A: Segments {s_i} are one-hot encoded, each representing a distinct anatomical region. We aggregate them by summing {s_i} to create composite mask with coarse anatomical structure. The aggregated mask is then normalized to the range [0, 1] to produce final S_HR.
R3: Typo of L_slign_d, L_slign_s A: Thanks for pointing that out.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The rebuttal adequately addressed most of the concerns raised by the reviewers.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A