List of Papers Browse by Subject Areas Author List
Abstract
Neurodegenerative diseases exhibit substantial phenotypic heterogeneity arising from the interplay of genetic risk and non-genetic factors, which is often obscured in standard neuroimaging analyses. Although structural MRI captures brain atrophy, it reflects a final overlapping pathway of diverse etiological processes, making it difficult to disentangle genetically driven degeneration from environmental or lifestyle-related effects. We propose a Genetically Guided Autoencoder (GG-AE), a self-supervised framework that leverages disease-associated genetic variation as a biological anchor to constrain the representation learning of brain atrophy. The framework first learns a compact genetic latent representation using a variational autoencoder, and subsequently decomposes voxel-wise imaging data into a genetically aligned subspace and a residual subspace using a dual-latent imaging autoencoder with non-negative, part-based decoding. On semi-synthetic data with known ground truth, GG-AE achieves improved recovery of genetically driven atrophy patterns compared to unguided and correlation-based baselines. When applied to real ADNI data, the model identifies spatially interpretable atrophy patterns and genetic associations that are consistent with established Alzheimer’s disease findings, while isolating non-genetic variation in a residual component. These results demonstrate the potential of genetically guided representation learning for disentangling disease heterogeneity and enabling mechanistically grounded imaging–genetics analysis. Our code is available at https://github.com/marilenadp/GG-AE.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5853_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/marilenadp/GG-AE
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DeMar_GGAE_MICCAI2026,
author = { De Pian, Marilena AND Wen, Junhao AND Davatzikos, Christos},
title = { { GG-AE: Genetically-Guided Autoencoder for Disentangling Structural Networks of Brain Atrophy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a genetically guided autoencoder (GG‑AE), a self‑supervised framework for disentangling brain atrophy patterns into genetically driven components and non‑genetic residual components from structural MRI and SNP 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.
1.It adopts a genetically guided paradigm instead of simple multimodal fusion, making the modeling strategy more consistent with the biological direction from genotype to phenotype. 2.It introduces non-negativity and soft orthogonality constraints to derive additive, localized brain structural patterns, leading to strong interpretability.
- 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 abstract and introduction overstate the contributions by claiming to be “the first to disentangle genetically driven versus non-genetic brain atrophy”, despite the existence of prior work addressing similar problems. 2.The model demonstrates insufficient novelty and is merely a trivial combination of existing mature techniques. It integrates VAE, dual-latent autoencoder, NMF-style decoder, and InfoNCE alignment, none of which are new. No novel network architecture, loss function, or disentanglement theory is proposed. The so-called “genetic guidance” is essentially contrastive alignment plus orthogonality regularization, which has been extensively explored in multimodal disentanglement research, leading to no theoretical contribution. 4.The paper does not compare against state-of-the-art multimodal disentanglement methods such as disentangled VAE, MoCo, DIP, and other modern contrastive disentanglement frameworks. It only evaluates against conventional baselines including NMF, vanilla autoencoder, and Deep CCA. Ablation analyses are seriously incomplete. While the paper includes permutation tests and a Deep CCA variant, it lacks systematic ablation on critical components: non-negativity constraints, soft orthogonality constraints, dual-latent structure, and two-stage genetic pre-training. It remains unclear whether each design choice is independently essential or only provides marginal improvement via simple component stacking.
- 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 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.
(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?
The score assigned to this paper is primarily determined by its major weaknesses, which have been detailed in the box above.
- 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 author’s explanation regarding methodological innovations did not convince me. I stand by my original decision.
Review #2
- Please describe the contribution of the paper
The authors propose a staged autoencoder framework (GG-AE) that first leverages a variational autoencoder (VAE) to learn a representation of genetic risk from SNPs and an imaging autoencoder with a Non-Negative Matrix Factorization (NMF)-inspired decoder for interpretability through additive, part-based spatial maps. The imaging autoencoder’s latent space is split into a subspace aligned with genetic representation from the frozen vae and the residual subspace is for non-genetic variation.
They evaluate on semi-synthetic data that’s derived from the UK BioBank (UKB) and then also on real Alzheimer’s Disease Neuroimaging Initiative (ADNI) 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.
The NMF-inspired decoder is well-motivated and practical. This is because this approach produces addtive, spatially interpretable components.
Their overall approach of explicitly partitioning genetic and then non-genetic into a residual subspace is well-motivated/novel as most fusion approaches tend to treat them symmetrically. This is good because the authors thus allow their method to respect real-world directionality (genotype goes to some intermediate molecular proces which then results in the phenotype). A symmetric version would blend genetic/non-genetic and make it unclear which modality is actually contributing.
Their ADNI results look biologically meaningful/relevant. This is because the genetic patterns found line up with known AD neuroanatomy and known SNP associations, also I like that their residual part shows no significant genetic association.
- 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.
My primary concern is in regard to using 54 pre-selected AD-associated SNPs as input to the real-data evaluation. With all input SNPs already curated for AD relevance, the VAE’s representation learning task is pretty simplified, since it’s compressing 54 AD signals into K dimensions rather than distinguishing AD-relevant from AD-irrelevant genetic variation. This means the downstream imaging model is constrained to align with an already AD-enriched latent space, which, in my opinion, stacks the deck in favor of recovering AD-relevant spatial patterns. To be clear, I’d say this does not predetermine the spatial maps or the partitioning across K=3 components, but it does make one wonder how dependent the framework’s performance is on this curated input. Since this paper presents GG-AE as a general framework, but the real-data evaluation relies on a highly curated SNP set that eliminates the challenge of distinguishing disease-relevant from irrelevant genetic variation, a task the VAE would need to handle for less well-characterized diseases or in discovery-oriented settings. I’d say it remains unclear how sensitive the framework is to the composition and purity of the input SNP set.
QUESTION: Can the authors report/comment on the expected sensitivity of the framework to SNP set composition?
I do note that the semi-synthetic experiments partially address this concern with the ADNI pre-selected SNPs, and I appreciate that it’s grounded in actual/real UKB data. However, the genetic simulation lacks features that would make it a convincing test. Specifically, the 95 non-risk SNPs are drawn independently from baseline MAFs with no correlation structure, making the 5 causal variants per pattern trivially identifiable. In real genotype data, risk SNPs are embedded in haplotype blocks where neighboring variants carry partial signal through LD, and polygenic background contributes correlated noise across many loci. Additionally, there is no pleiotropy (each SNP maps to exactly one pattern) and no population stratification. Thus, I’m still unsure regarding input sensitivity to realistic/real-world SNP input/set composition.
- 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.
(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?
I think the authors core proposed approach (asymmetry for genetic guidance with NMF-inspired decoding for interpretability) is well-founded. ADNI results themselves look solid biologically. Overall methodological design is generally robust, I appreciated the Perm GG-AE experiment for architectural bias as an example.
As outlined in the weakness section, I find the SNPs input to the real-world ADNI test data experiment to be potentially biased, thus making it unclear how much of this is actually the GG-AE itself learning meaningful structure in real data. While I appreciate the semi-synthetic grounded in UKB, I note my weakness section comment regarding the simulation setup.
As such, I’m giving a weak accept because the core approach is well-motivated with reasonable results, but I’d like greater clarification regarding SNP input set impact/sensitivity to the GG-AE.
- 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.
I vouch for the paper to be accepted. The authors clarify in their rebuttal that GG-AE is meant to take in some already curated SNP set to guide the learned imaging representation/spatial patterns.
For example, this could be from existing biological knowledge or other data-driven methods explicitly developed for SNP-disease association discovery (where there are numerous methods already developed).
As such, the core novelty that asymmetry in their design allows for voxel-imaging dimensionality to not drive the learned imaging-genetic representations, but instead allow underlying genetic information to be the driver is a useful methodological contribution that could be applied to other disease with well characterized SNP sets and varying structural/spatial patterns in imaging data.
Review #3
- Please describe the contribution of the paper
The study introduced genetically guided autoencoder (GG-AE) , a self-supervised approach that guides the learning of brain atrophy representations by using disease-relevant genetic variation as a biological reference, and showed that the resulting atrophy patterns and genetic relationships were spatially interpretable and aligned with prior Alzheimer’s disease knowledge.
- 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 strength of this paper lies in the fact that it first validates the feasibility and validity of the proposed model using synthetic data, and further demonstrates, through real ADNI data, that Alzheimer’s disease related SNPs emerge meaningfully, thereby supporting both the biological relevance of the method and its applicability to real data.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
This study is limited in that it adopts a relatively simple genetic variant reconstruction strategy for disentanglement in the process of genetic representation learning. As a result, it has not been sufficiently validated how precisely the latent representation disentangles and structures genetic factors.
- 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?
It would have been beneficial to include an additional loss term that explicitly encourages the learned representations to preserve genetic characteristics, which could further support the validity of the genetic representation learning framework.
- 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 have clarified the rationale for using a VAE to embed SNPs in a clearer and more understandable manner.
Author Feedback
We thank the reviewers for their valuable feedback. R1 appears to interpret our paper as a general-purpose disentanglement framework, leading to critiques regarding limited novelty in disentanglement theory, lack of comparisons with broad multimodal disentanglement models, and reliance on existing components. However, as also noted by R2/R3, this work is primarily a biologically motivated model for interpretable imaging-genetics discovery rather than a general disentanglement theory contribution. Specifically:
1.Interpretable structural pattern discovery remains underexplored in recent deep imaging-genetics work, which has focused largely on classification. Unlike most prior approaches, GG-AE is an unlabeled, data-driven framework for mechanistic representation learning without disease labels.
2.A key contribution is integrating an NMF-inspired decoder within a deep autoencoder. Although widely used for interpretable neuroimaging decomposition, NMF is inherently linear with dedicated optimization procedures. GG-AE preserves NMF-style interpretability while enabling scalable self-supervised learning with more expressive encoders. So, we integrated contrastive alignment and orthogonality regularization within an interpretable NMF-based imaging-genetics framework to disentangle NMF components, which to our knowledge has not been explored previously.
3.Voxel-level imaging creates severe dimensional imbalance that can dominate genetic signals. GG-AE, instead of symmetric fusion, directly models the biological direction from genotype->phenotype by using genetic variation to guide imaging representations, as also recognized by R2. Thus, while GG-AE combines existing methodological components, our contribution lies in their biologically motivated integration for interpretable self-supervised imaging-genetics analysis.
Regarding baselines/ablations (R1), we focused on methods commonly used in unsupervised imaging-genetics analysis, since the goal was mechanistic representation learning rather than benchmarking all multimodal disentanglement models. Existing disentanglement methods are rarely evaluated for self-supervised imaging-genetics settings with voxel-wise neuroimaging and severe modality imbalance, and implementing them was beyond this study’s scope. We therefore used representative baselines including Deep CCA, imaging-only decomposition methods, and permutation controls. While additional ablations could further isolate individual design choices, the 8-page limit led us to prioritize controlled semi-synthetic validation with known ground truth and application to real data, as appreciated by R2/R3. Regarding SNP selection (R2) and genetic VAE modeling (R3), GG-AE is intentionally guided by the selected SNPs relevant to the biological question, and the current real-data experiment therefore focused on AD-related variants. Our goal was not to distinguish AD- from non-AD-related SNPs or learn a fully disentangled genetic factorization, but rather to learn a compact representation that guides separation of heterogeneous structural atrophy patterns within the AD spectrum. SNP selection can be adapted to different objectives, including disease-specific, multi-disease, or LD-pruned variant sets. Because SNP inputs are sparse we used a VAE to capture latent genetic distributions rather than deterministic embeddings alone, while progressively increasing β-weighted KL regularization and sigmoid-transformed latent sampling encouraged stable distributed representations. The genotype simulation was designed as a controlled proof-of-concept to evaluate if GG-AE can recover genetically associated structural components while separating residual variation, rather than model full genetic architecture, including LD and polygenic effects. Nevertheless, the observed SNP-latent associations, imaging alignment behavior, and permutation-control experiments support the biological relevance of the learned representations.
Code will be released.
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 reviews were mixed, with R2 and R3 recommending weak accept but R1 recommending reject. All reviewers raised concerns about the paper that should be addressed in the rebuttal. Besides the other mentioned fixable and addressable points answer the novelty and contribution concerns of R1 should be part of the rebuttal.
- 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.
Following the rebuttal, two of the three reviewers recommended acceptance. R2 and R3 found the clarifications regarding the paper’s positioning as a biologically motivated imaging-genetics framework rather than a general-purpose disentanglement contribution convincing. R1’s concerns about missing comparisons and incomplete ablations reflect a different evaluation perspective rather than fundamental technical issues and are offset to a reasonable extent by the controlled, semi-synthetic validation and the biological plausibility of the ADNI findings. The final version should moderate the novelty claims and explicitly acknowledge the scope of the ablation study as a limitation.
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.
Based on the rebuttal and review comments, reviewers seem to be satisfied by the provided explanation save the claims on methodological novelty (which I agree with to an extent). Since the strengths of the contribution appear to outweigh the demerits, I vote for accepting this work
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After considering the reviews and rebuttal, I recommend acceptance. The paper proposes a genetically guided autoencoder for interpretable imaging-genetics analysis of brain atrophy, combining genetic guidance with an NMF-inspired decoder to separate genetically associated structural patterns from residual variation.
The rebuttal clarifies that the contribution is not intended as a general disentanglement theory, but as a biologically motivated framework for interpretable imaging-genetics discovery using curated SNP sets. This addresses the main concern about task positioning. The semi-synthetic validation and ADNI results provide reasonable evidence that the learned patterns are biologically meaningful. Some limitations remain, including incremental use of existing components, limited ablation against broader disentanglement methods, and sensitivity to SNP-set selection.