Abstract

Magnetic resonance imaging (MRI) provides detailed soft-tissue characteristics that assist in disease diagnosis and screening. However, the accuracy of clinical practice is often hindered by missing or unusable slices due to various factors. Volumetric MRI synthesis methods have been developed to address this issue by imputing missing slices from available ones. The inherent 3D nature of volumetric MRI data, such as cardiac magnetic resonance (CMR), poses significant challenges for missing slice imputation approaches, including (1) the difficulty of modeling local inter-slice correlations and dependencies of volumetric slices, and (2) the limited exploration of crucial 3D spatial information and global context. In this study, to mitigate these issues, we present \textbf{S}patial-\textbf{A}ware \textbf{G}raph \textbf{C}ompletion \textbf{N}etwork \textbf{(SAGCNet)} to overcome the dependency on complete volumetric data, featuring two main innovations: (1) a volumetric slice graph completion module that incorporates the inter-slice relationships into a graph structure, and (2) a volumetric spatial adapter component that enables our model to effectively capture and utilize various forms of 3D spatial context. Extensive experiments on cardiac MRI datasets demonstrate that SAGCNet is capable of synthesizing absent CMR slices, outperforming competitive state-of-the-art MRI synthesis methods both quantitatively and qualitatively. Notably, our model maintains superior performance even with limited slice data.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0554_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{LiuJun_SAGCNet_MICCAI2025,
        author = { Liu, Junkai and Aung, Nay and Arvanitis, Theodoros N. and Piechnik, Stefan K. and Lima, Joao A. C. and Petersen, Steffen E. and Zhang, Le},
        title = { { SAGCNet: Spatial-Aware Graph Completion Network for Missing Slice Imputation in Population CMR Imaging } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {460 -- 469}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposed a missing slice imputation framework, SAGCNet, that utilize a newly proposed volumetric slice graph completion (VSGC) which models the slices as graph vertices and try to learn the inner-slice relationship. The volumetric spatial adapter (VSA) modules is utilized as the image encoder. Experiments on three datasets are performed, and both quantitative and qualitative results are reported and SAGNCNet achieves the highest performance in all cases.

  • 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 idea of introducing graph to improve the learning of inner-slice relationship is innovative. The modules are carefully designed and the writing is clear.

  • 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.
    • Unclear intuition - while utilizing graph is a reasonable choice, but the authors don’t discussion what’s the advantage of the graph approach versus 3D vision transformer, like UNetR, which can also effectively exploit long-term dependency among the slices
    • Limited performance improvement and potential heavier computational cost. Telling from Table 1, the proposed method usually have relatively minor performance improvement compared with UNetR, and often smaller than 0.1 with larger slice missing rate. The limited performance is hard to justify its potential increase in the computational cost and hyperparameter tuning. Besides, telling from table 2, it seems the VSA module, which is not new, contributes to a large portion of the performance improvement. This further damage the significance of the graph module.
    • Problematic evaluation - from the visual comparison it seems the authors didn’t perform any cropping to remove the background, and from the error maps it seems a lot of errors are from the background (non-cardiac) region, which is typically excluded from evaluation since they don’t provide diagnostic information and this also decreases the significance of the improved performance.
  • 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 provide sufficient information for 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.

    (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 idea of using graph to model the inner-slice relationship is novel and refreshing, but the unclear intuition of the advantage over 3D ViT and the limited performance with extra computation al cost decrease the significance of this paper.

  • 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.

    Thanks to the authors’ clarification on my questions, but the clarification still cannot convince me:

    1. Justification for using graph representation over 3D ViT/ UNETR alternatives: the authors explain that the proposed method can capture non-Euclidean relationship while the baselines cannot. However, the attention mechanism in the transformer baselines can also capture such relationship. What’s more, for the particular application for missing slice synthesis, it’s not very clear why non-Euclidean approaches would be significantly better than Euclidean approaches, since intuitively the missing slice should be mostly relevant to the Euclidean neighbor slices.
    2. Concern on the performance: the authors stresses the difficulty of the tasks and the proposed method outperforms all baselines, which are both true. However, the clarification does not address the concern on the minor performance gain on PSNR / SSIM, which is about 1% improvement (e.g. PSNR=19.39 vs 19.31) in most cases and 4.5% (PSNR=38.15 vs 36.55) in only one case, according to table 1.
    3. Concern on computational cost: architecture-wise, the proposed methods add additional graph modules to the UNetR architecture. Given the limited performance gain as mentioned above, while the additional module can be lightweight and fast, still the additional computational cost brings more efficiency sacrifice than performance gain
    4. Concern on using the whole MRI scan, including non-heart region, instead of ROI at the heart region: Indeed I cannot deny the background region can potentially be very useful. But for most cardiac research works, people only focus on the heart region, sometimes even only the left ventricle region. So it would be great to include the evaluation on both heart region and whole-image, but having only whole-image evaluation greatly decrease the clinical significance in many applications, especially from figure 4 we can clearly see a lot of error of the baseline UNetR is from the out-of-heart background region.

    The proposed work is indeed quite inspiring, but given the insufficient justification, minor performance gain and improper evaluation on whole image instead of heart ROI, it needs more polish and improvement to be accepted.



Review #2

  • Please describe the contribution of the paper

    The authors proposed a novel Graph neural network-based model for CMR missing slice imputation. In the proposed model, the CMR scan with missing slices is represented as a graph and completed by two graph attention networks. To preserve 3D spatial information, the authors introduced an efficient conv-based module in the transformer encoder.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1) The paper is well-written and easy to follow.

    2) The idea of modelling the incomplete MRI as a graph is novel and interesting, and the experiment shows promising results.

  • 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 author did not provide enough information for the graph construction. How does it work exactly? Would missing slices (their initial features should be very similar) all be connected by using KNN? Would it be easier to construct the graph with the adjacent slices? I can imagine that this neighbourhood-based construction might result in an acyclic graph, which makes more intuitive sense compared to the cycles produced by KNN (Fig.3). I would then argue that perhaps a graph is not suitable for representing a set of consecutive images, which is not a ‘cycle’.

    2) More clarification is needed for better reproducibility. This includes but is not limited to: What’s the input of the model? Only the available slices or the entire 3D volume with missing slices zero-padded? What kind of perceptual loss was used?.

  • 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.

  • 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.

    (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?

    This paper introduces a novel and interesting idea for missing slices imputation. The experiments also show promising results. However, the authors should further clarify why the graph is suitable for representing CMR slices.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors answered my questions. Although there are still some open questions left, I think this paper presents an intereting idea and is worth further discussion.



Review #3

  • Please describe the contribution of the paper

    This paper addresses the challenge of missing slice imputation in cardiac MRI volumes, a common clinical issue caused by factors like motion artifacts and scanning time constraints. The authors introduce SAGCNet (Spatial-Aware Graph Completion Network), which combines two novel components: a graph-based structure that models relationships between slices, and spatial adapters that capture 3D contextual information. The method treats MRI slices as nodes in a graph and uses a multi-view approach to handle missing data at both attribute and structure levels. Experiments across three cardiac MRI datasets show that SAGCNet outperforms existing image synthesis methods, particularly in preserving anatomical details. The ablation studies confirm that both the graph structure and spatial adapters contribute meaningfully to the model’s effectiveness at synthesizing missing slices from incomplete volumetric data.

  • 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.

    [Introduction]

    • “Nonetheless, missing slice is a common issue for volumetric MRI data, particularly in cardiac magnetic resonance (CMR) volumes, within clinical applications and practice, caused by factors such as excessive scanning times, image deterioration, motion artifacts, and disparate acquisition techniques [2]. Consequently, the development of a unified and effective approach for imputing missing slices using available data is critically needed [27].” The authors provide a clear and well-articulated interpretation of the fundamental challenges associated with CMR image volumes in clinical settings.

    • “Below, we highlight the following two key challenges that need be addressed for missing slice imputation task:” The authors present a precise and incisive characterization of the challenges involved in the missing slice imputation task, which effectively frames the subsequent methodological contributions.

    [Section 1: Introduction - Contributions]

    • “We propose a volumetric slice graph completion (VSGC) module, employing graphs to capture inter-slice correlations. To the best of our knowledge, our work is the first to leverage graph-based modeling at slice level for medical image synthesis.” The conceptualization of modeling slices as graph nodes represents a novel approach to capturing inter-slice relationships, demonstrating innovative thinking in addressing the missing slice imputation problem.

    [Section 3: Experiments]

    • “3 Experiments” The experimental validation against multiple baseline methods using standard evaluation metrics is commendable, although a more detailed ablation study and thorough analysis of each component’s contribution would strengthen the validity of the proposed methodology.
  • 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.

    [Section 2: Methodology]

    • “2 Methodology” The methodology appears promising, though concerns arise regarding whether SAGCNet requires complete 3D volumes as ground truth for training. Given the inherent problem of slice availability discussed in the introduction, and considering datasets like ACDC with 5-8mm slice spacing (implying inherent missing inter-slice dependencies), clarification on how SAGCNet addresses these issues would be valuable. Additionally, elaboration on how the model handles slice misalignment problems common in CMR images would enhance the discussion.

    • “Next, the k-Nearest Neighbors (kNN) algorithm is used to construct the incomplete volume slice graph G = {V, E}, where V and E represent the node (slice) and edge (inter-slice correlation) set, respectively.” Additional clarification regarding the mechanism by which the kNN algorithm transforms volumetric data into graph structures would enhance understanding of this crucial step in the methodology.

    • “Notably, nodes corresponding to missing slices are treated as nodes with missing attributes, and their associated edges are also considered incomplete.” Further elaboration on the specific implementation of nodes with incomplete edges would be beneficial for a comprehensive understanding of the graph representation.

    • “Regarding structure-view completion, we utilize personalized PageRank [9] to propagate information and enhance its diversity, producing Gs = (Xs, As), which is crucial for learning comprehensive and discriminative node representations under conditions where partial edges are missing.” The intuition behind utilizing PageRank for modeling node attributes and graph relations despite missing attributes remains insufficiently explained. A more detailed exposition of this mechanism would strengthen the methodological foundation.

    • Lrec, Lsyn consist of L1 loss and perceptual loss.” Identification of where these two losses are computed within the architecture depicted in Figure 2, along with specific computation details, would improve clarity regarding the training process.

    [Figure 3]

    • “Fig. 3. An illustration of the VSGC module.” Clarification of what the notation ‘C’ represents in this figure would be helpful, particularly whether it corresponds to the number of available slices as hypothesized.

    [Section 3: Experiments]

    • “Zero-padding is applied for all volumes in the through-plane direction to ensure the fixed input size.” Additional information regarding the specific fixed size used in the through-plane direction (presumably N as introduced in the methodology section) and the rationale for its determination would enhance reproducibility.

    • “Table 1. Quantitative performance comparison of different methods on three datasets under three missing rate configurations. The best performance is in bold.” The marginal improvement in evaluation metrics compared to the UNETR backbone warrants deeper discussion. A more comprehensive evaluation from multiple perspectives would provide better insights into the functional contributions of the proposed components.

    • “Fig. 4. Qualitative results of all methods on the UKBB dataset. Every two rows, from top to bottom, denote the experimental results and the error maps, respectively.” The visualization would benefit from inclusion of a scale to indicate high and low error values in the error maps, enhancing interpretability of the qualitative results.

    • “Table 3. Quantitative performance comparison of ablation variants.” The relatively modest performance differences when removing key components (VSA or VSGC) from the complete network raises questions about their individual contributions. A more in-depth analysis of component functionality would strengthen the justification for the proposed architecture.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (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?

    I recommend accepting this paper based on several notable strengths. First, the authors demonstrate excellent problem formulation by specifically targeting the clinically relevant issue of missing slices in CMR volumes, providing a clear motivation for their work. Second, the methodological approach is well-conceived, with the innovative representation of slices as graph nodes and the incorporation of spatial adapters to capture 3D contextual information—both backed by sound theoretical assumptions regarding inter-slice relations. Third, the experimental validation is comprehensive, featuring comparisons against multiple state-of-the-art baselines across three datasets, with results consistently demonstrating SAGCNet’s superior performance in both quantitative metrics and qualitative visual quality. While there remain opportunities for more in-depth analysis of component contributions, the paper makes a valuable contribution to the field of medical image synthesis by introducing a novel graph-based approach to missing slice imputation.

  • 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.

    Recommendation: Accept

    This paper presents a novel graph-based approach to a clinically relevant problem in cardiac MRI, demonstrating substantial validation advances across multiple datasets against strong baselines. The authors’ rebuttal effectively addressed my concerns regarding methodological details and implementation specifics, reinforcing the technical soundness of their contribution. The combination of clinical relevance, robust experimental validation, and comprehensive response to reviewer feedback supports acceptance.




Author Feedback

We thank all reviewers for their valuable and constructive feedback. We are encouraged that all (R1, R2, R3) found our graph-based modeling of slice relationships novel and clinically relevant, and the paper is clearly written and well structured. Below we address key concerns and clarify misunderstandings where appropriate.

(1) R1 & R3 – Justification for using graph representation over 3D ViT/ UNETR alternatives.

We appreciate the concern regarding the novelty and comparative value of our graph-based approach over 3D ViTs like UNETR. Unlike 3D ViTs/ UNETR, which rely on fixed Euclidean structure, our method captures non-Euclidean inter-slice relations, better handling anatomical irregularities. GNNs also support variable-length, sparse connections, complementing ViT’s dense attention. Furthermore, SAGCNet consistently outperforms UNETR, especially under sparse data, showing that explicit graph-based inter-slice modeling improves robustness and synthesis accuracy.

(2) R1 – Limited performance gain and computational cost.

While gains over UNETR may appear modest at high missing rates, SAGCNet consistently outperforms baselines across all datasets, demonstrating strong generalizability. More substantial gains at low missing rates (clinical practice) — highlight its real-world utility. The task itself is highly challenging, requiring both inter-slice dependency modeling and 3D spatial coherence, which most baselines fail to jointly address. Regarding computational cost, SAGCNet offers a favorable trade-off: similar parameter count to UNETR, faster convergence, and efficient design via lightweight VSA modules and sparsely constructed VSGC graphs.

(3) R1 – VSA contributes more than VSGC, limiting the graph module’s impact.

SAGCNet enhances UNETR with VSA and VSGC, addressing complementary challenges in slice imputation. VSA improves intra-slice spatial awareness, while VSGC models inter-slice relationships—crucial for anatomical continuity when slices are missing. Although VSA shows higher ablation gains, VSGC consistently contributes to performance and structural coherence in qualitative results. Their synergy enables SAGCNet to produce anatomically faithful and globally consistent generation.

(4) R1 – Evaluation concerns: possible background influence on error maps.

Including background ensures structural realism for downstream registration and segmentation tasks. Background carries cues aiding volumetric coherence which excluding it may hurt generalization. Evaluating on full images increases difficulty, SAGCNet still reduces error across both cardiac and background regions, showing robust global modeling.

(5) R2 & R3 – Clarifications on graph construction & missing slice representation.

Each node in the graph corresponds to a slice embedding from the encoder, where features from transformer blocks are aggregated via a channel adapter and flattened into a single vector. Edges are built using a top-k nearest neighbor (k=3) strategy based on feature similarity in latent space. This captures both adjacent and non-contiguous inter-slice relationships. While adjacent slices often show the highest similarity, kNN allows more flexible connectivity, especially when adjacent slices are missing, keeping the graph informative under sparse conditions. Although cycles may emerge, they do not harm semantic consistency and instead support robust information flow. In practice, kNN and adjacency-based graphs yield largely overlapping edges due to typical cardiac MRI slice spacing. We will clarify these design choices in Section 2.3.

(6) R2 & R3 – Model input and perceptual loss details.

SAGCNet takes zero-padded volumes as input to simulate real-world settings. The perceptual loss is computed using a pretrained VGG19 network, applied to synthetic and GT slices. This information will be made explicit in revised manuscript. We also commit to releasing our code and preprocessed datasets upon acceptance to enhance reproducibility.




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’

    N/A



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



back to top