List of Papers Browse by Subject Areas Author List
Abstract
Multimodal neuroimaging provides complementary information for early computer-aided diagnosis of Alzheimer’s disease (AD). However, in clinical practice, modalities such as PET are often missing due to cost and accessibility constraints, hindering stable cross-modal inference. Existing approaches either synthesize the missing modality in the image domain, which may introduce artifacts and noise that interfere with classification, or perform feature-domain alignment and completion while emphasizing spatial consistency, making it difficult to handle MRI-PET heterogeneity in texture patterns and frequency components. Moreover, forcing unpredictable cross-modal components to fit can inject noise and degrade diagnosis. To address these issues, this paper proposes the Spatio-Spectral Guided Affinity Imputation Network (S²GAIN), which employs a cascaded dual-stream cross-modal interaction backbone and multi-stage lightweight imputers to infer missing PET features layer by layer. A spatio-spectral joint alignment strategy is introduced to apply contrastive constraints in both the spatial domain and the frequency-domain magnitude spectrum, improving heterogeneous representation alignment. An affinity-guided reliability-weighting mechanism is further designed to suppress noise propagation from uninferable components. Finally, image-domain reconstruction regularization is employed to improve consistency between imputed metabolic representations and PET. Experimental results on ADNI demonstrate that S²GAIN achieves superior and consistent performance under missing-modality settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5203_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{WuYan_S2GAIN_MICCAI2026,
author = { Wu, Yanan AND Dong, Yihong AND Wu, Yuehan},
title = { { S2GAIN: Spatio-Spectral Guided Affinity Imputation Network for Alzheimer’s Disease Diagnosis with Incomplete Modality } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes S2GAIN that infers missing PET features layer by layer.
- 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 overall presentation of this paper is clear. 2.Missing PET is a crucial task in AD diagnosis. 3.Experiments are reasonable and presented clearly.
- 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.Ablation studies should not be limited to lambda, alpha, and beta. 2.On page 7, 1 line and Table 2 are not aligned nicely.
- 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?
This paper address a crucial task in AD diagnosis, and experiments are clear and justify the proposed model.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The main contributions of this study are as follows. First, to address missing PET data in Alzheimer’s disease diagnosis, the authors propose S²GAIN, a novel incomplete multimodal learning framework that progressively recovers PET representations from MRI in feature space, enabling stable diagnosis when PET is unavailable at inference. Second, they introduce S²JA, a spatio-spectral joint alignment strategy that improves cross-modal representation alignment by enforcing consistency in both the spatial and frequency domains. Third, an affinity-guided reliability weighting mechanism, AGRW, is developed to emphasize reliably predictable components while suppressing non-inferable noise from modality heterogeneity, thereby reducing noise propagation and enhancing robustness. Finally, the method incorporates image-domain reconstruction regularization via a PET reconstruction branch to preserve spatial and metabolic texture consistency of deep metabolic features, making the completed representations both diagnostically informative and structurally consistent with real PET.
- 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 methodological innovation is well targeted. The authors identify two critical limitations in existing missing-modality approaches and design corresponding solutions accordingly. First, most feature completion methods focus only on spatial-domain alignment, even though the discrepancy between MRI and PET lies not only in spatial structure but also in texture patterns and frequency components. Second, many studies implicitly assume that MRI can recover the full spectrum of PET information, whereas part of the metabolic information in PET is not reliably predictable from MRI; forcing such components to be fitted may introduce noise and further blur the classification boundary. To address these issues, the authors propose S2JA and AGRW. S2JA extends alignment from the spatial domain to a joint spatial-frequency framework by incorporating amplitude-spectrum constraints, while AGRW uses affinity to estimate the reliability of cross-modal completion, emphasizing more trustworthy correspondences and suppressing noise from inherently unpredictable components. The main novelty lies not merely in the use of FFT or weighting, but in a more refined modeling perspective: distinguishing how modality differences should be aligned and which components should not be forced to be learned. 2.The experimental evaluation is convincing, particularly the cross-cohort validation. This is another major strength of the study. Many related works rely solely on splits within a single dataset, which may yield strong performance but provide limited evidence of robustness under distribution shift. In contrast, this study conducts a bidirectional cross-cohort evaluation on ADNI-1 and ADNI-2, training on one cohort and testing on the other, and then reversing the setting. This design is substantially more rigorous than conventional random splits and offers stronger evidence of generalization. The stronger experimental protocol, together with results that closely match the methodological motivation, makes the overall empirical support more compelling.
- 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 applicability of the proposed method remains relatively narrow. Although the title claims to address the general problem of incomplete modalities, the study focuses primarily on a single missing-pattern setting, namely the availability of MRI with missing PET, and all experiments are conducted under this assumption. While this scenario is clinically meaningful, the paper does not establish whether the method would remain effective in more complex incomplete multimodal settings, such as simultaneous absence of multiple modalities or varying missing rates. The authors do acknowledge in the conclusion that future work should extend the framework to more challenging missing-modality scenarios. 2.The reliability weighting mechanism still depends on supervision from real PET data. The idea of AGRW is reasonable, yet during training it requires the affinity between completed features and ground-truth PET features to generate reliability weights. In essence, this means that the mechanism still relies on the availability of true PET supervision at the training stage. Therefore, although effective under the current experimental setting, the paper does not clarify whether this module would retain the same advantage in scenarios where paired PET data are scarcer and supervision is weaker. 3.The validation of completion quality is still not sufficiently direct. A central goal of this study is to recover PET representations from MRI, and the authors further introduce image-domain reconstruction regularization to encourage the completed representations to better approximate the metabolic distribution and structural consistency of real PET images. However, the experiments mainly report downstream classification results, together with ablation studies and parameter sensitivity analyses, without providing more direct evidence of completion quality, such as quantitative similarity metrics between completed and real PET or visual comparisons of representative cases. As a result, the current findings demonstrate more clearly that the method benefits classification, while offering less direct support for the claim that it produces more faithful PET representations. 4.Although the paper compares the proposed method with several advanced missing-modality imputation approaches, it omits a particularly important baseline: a strong MRI-only classifier. Without reporting the performance of a model trained directly on MRI alone, using either a standard backbone such as ResNet or the paper’s own 3D Dense Block, it is difficult for readers to assess the absolute performance gain brought by the proposed cross-modal completion framework over a single-modality MRI setting. 5.Some writing issues are also present in the manuscript. On page 3, in Contribution 1, there is a missing space before “spectrum”. In Contribution 2, the ending period is missing. On page 6, after DfpetL, there should be a space following the period.
- 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.
(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 primary reason for my Weak Accept recommendation is that this work proposes a relatively comprehensive and well-targeted solution to the practically important problem of missing PET data in Alzheimer’s disease diagnosis. Rather than simply combining existing modules, the method is designed to address two critical limitations of current missing-modality approaches. First, it extends cross-modal alignment from the spatial domain alone to a joint spatial-frequency modeling framework. Second, it introduces a reliability weighting mechanism to explicitly suppress noise propagation caused by components that are inherently difficult to infer from MRI. These designs show clear novelty and are well aligned with the problem setting. Furthermore, the paper adopts a bidirectional cross-cohort evaluation between ADNI-1 and ADNI-2.Compared with the common practice of using only within-dataset train/test splits, this setting provides stronger evidence of generalizability, making the experimental protocol itself more convincing.
However, I do not think the current evidence is sufficient to justify a stronger recommendation. First, the method is mainly validated under a single missing-modality pattern, namely MRI available and PET missing. Although the title suggests a broader incomplete-modality setting, the manuscript does not yet demonstrate the effectiveness of the method in more complex missing scenarios. Second, AGRW still relies on ground-truth PET supervision during training, and the paper does not clarify whether this mechanism would remain effective when paired PET data are more limited. Third, although the paper emphasizes that the proposed method can recover representations that are closer to real PET, the experiments mainly report downstream classification performance. More direct evidence on the quality of modality completion is still missing, such as similarity analysis against ground-truth PET, representative visualizations, or comparison with a stronger MRI-only baseline. As a result, while the paper provides reasonably strong evidence that the proposed method improves classification performance, it offers less direct empirical support for the claim that it more faithfully recovers PET representations. In addition, the manuscript contains several minor writing and formatting issues, suggesting that the overall presentation could be further improved. Overall, I believe this work shows promising novelty and potential practical value, and it is worthy of publication. However, since the experimental validation and supporting arguments are not yet fully comprehensive, Weak Accept is the most appropriate recommendation at this stage.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes Spatio-Spectral Guided Affinity Imputation Network (S2GAIN) for early computer aided diagnosis of Alzheimer’s disease. Also This paper enhances heterogeneous representation alignment by jointly exploiting spatial domain information and the frequency domain magnitude spectrum.
- 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.
A major strength of the paper is that it explicitly addresses the inherent heterogeneity between MRI and PET when synthesizing missing modalities, including differences in texture patterns, frequency characteristics, artifacts, and noise. Rather than relying solely on image-space information, the method jointly leverages spatial-domain and frequency-domain information, which is a well-motivated design for improving heterogeneous representation alignment across modalities. This design is meaningful because multimodal AD diagnosis often depends on effectively integrating complementary MRI and PET information under missing-modality settings.
- 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.
One important limitation is that the paper mainly evaluates the generated features through diagnostic performance, without sufficiently examining their interpretability or modality-specific validity. As a result, it remains uncertain whether the generated representations faithfully capture the characteristics of the target modality or simply act as useful auxiliary signals for prediction.
- 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 idea of jointly leveraging spatial domain and frequency domain information is well motivated. However, the paper should provide evidence regarding whether the generated features are truly meaningful and reliable.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Reviewer 1.Thank you for the suggestion. We selected λ1, λ2, λ3, α, and β for sensitivity analysis because they directly correspond to the three core objectives of S²GAIN, i.e., S²JA, AGRW, and R², and control the balance between spatial and spectral alignment. Thus, they are the most relevant objective-level hyperparameters. We agree that other factors, such as temperature τ, stage number L, imputer capacity, and PET decoder configuration, are also worth studying. However, they are mostly internal or structural hyperparameters, and changing them would affect model capacity, computation, and memory simultaneously. Due to the page limit, we prioritized the loss-related parameters and will study more settings in future extended work. We also thank the reviewer for pointing out the layout issue of Table 2. Reviewer 2.Thank you for the comments. We agree that this work focuses on the clinically common setting where MRI is available but PET is missing. MRI is generally more accessible and less costly, whereas PET is more likely to be absent due to cost, equipment requirements, and radioactive tracers. Thus, our goal is to impute PET-related latent representations from MRI for PET-missing AD diagnosis. More complex cases, such as multiple missing modalities, different missing patterns, or varying missing rates, are not covered here. We will explore them in future extended work, e.g., missing both fMRI and PET after introducing additional modalities. Regarding AGRW, we agree that it uses real PET features during training. Our method follows a supervised cross-modal imputation setting: paired MRI-PET data are used to learn the MRI-to-PET mapping, while only MRI is required during PET-missing inference. AGRW does not remove PET supervision; instead, it uses the affinity between imputed and real PET features as a reliability weight to emphasize predictable cross-modal associations and suppress non-inferable components. Regarding completion quality, we agree that similarity metrics or visual comparisons would provide more direct evidence. However, our goal is not PET image synthesis for visual inspection, but PET-related feature imputation for diagnosis. Thus, the current evaluation focuses on diagnostic performance, cross-cohort generalization, and module ablation. R² is an auxiliary regularizer rather than a standalone synthesis module; Table 3 shows that removing R² decreases performance, suggesting its benefit for structural consistency and diagnostic usefulness. The current results mainly support diagnostic effectiveness rather than pixel-level PET recovery. We will study direct completion metrics and visual comparisons in future extended work. Regarding the MRI-only baseline, we agree that it is important. The current comparisons focus on incomplete-modality learning and imputation methods under a similar setting. Table 3 shows that removing S²JA, AGRW, or R² degrades performance, indicating that the gain is not solely from the backbone. Still, these ablations cannot replace a direct MRI-only baseline. Due to the page limit, this comparison is not included, and we will add it in future extended work. We also thank the reviewer for pointing out the writing issues.
Reviewer 3.Thank you for the important comment. We agree that diagnostic performance alone cannot fully demonstrate interpretability or modality-specific validity. In S²GAIN, S²JA constrains MRI/PET features in spatial and frequency domains; AGRW uses feature affinity to focus on PET-related components inferable from MRI; and R² encourages deep PET-related features to preserve image-domain spatial structure. These designs encourage PET-consistent representations, but do not fully establish interpretability or target-modality validity. The current experiments mainly show diagnostic usefulness, and we will further study interpretability and modality-specific validation in future extended work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
A well-motivated paper with consistenly recognized techinical contributions in PET data imputation for AD diagnosis. Deeper discussion, ablation, and validation of the proposed method could further enhance the paper’s quality.