List of Papers Browse by Subject Areas Author List
Abstract
Dementia screening remains a clinical challenge due to the high cost, invasiveness and sometime subjective of traditional neuroimaging or cognitive assessment questionnaires. While retinal imaging provides a non-invasive alternative, current methods struggle to capture the cerebral pathology needed for accurate diagnosis. To address this, we propose EyeDEM, a novel framework designed to distill brain structural priors into a standalone eye screening network using paired eye-brain data. To overcome cross-organ heterogeneity, we disentangle the latent space to isolate cross-modal relate representations and then apply multi-level alignment to map retinal microvasculature to brain atrophy. Furthermore, we introduce a divide-and-conquer distillation strategy that selectively transfers hierarchical diagnostic knowledge from a fused multimodal teacher to the eye student. At the inference stage, EyeDEM functions as a standalone screening method requiring solely retinal images. Validated across multi-center datasets of 2,040 participants, our standalone eye model reaches the AUC of 0.888 for dementia vs normal control classification. Crucially, we validate the model’s interpretability through biological correlation analysis, demonstrating that the learned retinal representations strongly correlate with critical neurodegenerative brain regions by predicting their volumes. These findings confirm that leveraging the eye-brain axis as a cross-modal prior paves the way for accessible, high-precision, and clinically transparent dementia screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0746_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{XuZhu_CrossModal_MICCAI2026,
author = { Xu, Zhuoting AND Cao, Le AND Zhang, Tiande AND Mou, Lei AND Zhou, Shijia AND Li, Fan AND Shen, Dinggang AND Fu, Huazhu AND Wu, Bo AND Zhao, Yitian},
title = { { Cross-Modal Alignment and Distillation for Retinal Imaging Dementia Screening } },
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 EyeDEM, an innovative cross-modal framework for dementia screening. It is designed to isolate cross-modal shared features from modality-specific features through latent space disentanglement and employs a multi-level alignment mechanism to capture the continuous trajectory of neurodegeneration. The motivation behind this design is sound, and the proposed method demonstrates strong performance in Dementia and Normal Control classification using solely retinal 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.
This paper proposes EyeDEM, an innovative cross-modal framework for dementia screening. It is designed for isolating cross-modal shared features from modality-specific features through latent space disentanglement, and employs a multi-level alignment mechanism to capture the continuous trajectory of neurodegeneration.
The motivation of this design is clear, the proposed method in evaluation shows strong performance for Dementia and Normal Control classification with sole retina 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.
1.Ambiguity in Gradient Flow and Feature Alignment. The authors claim that the instance-level alignment (Eq. 2) ensures $\mathcal{C}E^+$ learns cross-modal features without updating the teacher (brain) representations. However, the enforcement of this constraint is unclear. Since $\mathcal{M}_B^+$ is involved in the cross-attention mechanism and the subsequent loss $\mathcal{L}{ins}$, gradients will naturally backpropagate to the Brain Encoder unless an explicit stop_gradient operation is applied. Furthermore, in the “Optimization” section, the authors state that the “full framework” is optimized in the second stage. If the BrainPipe (pre-trained in Stage 1) remains trainable, $\mathcal{L}_{ins}$ will inevitably modify/affect the brain representations leaning to align with the retinal features, which contradicts the premise of maintaining a fixed teacher prior mentioned in this paper. The authors should clarify the exact gradient flow and whether the teacher model is strictly frozen.
2.Incomplete Mathematical Definitions and Notations. Several notations and loss components lack formal mathematical definitions, hindering the reproducibility and comprehensibility of the proposed mechanism. In Eq. (3), the “tilde” notation for the cross-modality related representation of Brain is entirely unexplained. In Eq. (6), the weak distillation loss ($\mathcal{L}{weak}$) is introduced without a mathematical formulation. Additionally, there is insufficient explanation regarding how the high-level features $H_E$ and $H{B+E}$ are derived and what specific semantic or structural information they encode.
3.Minibatch Variance in Pearson Correlation Loss Regarding the Pearson Correlation Loss for predicted MoCA scores, it appears the correlation is calculated at the minibatch level, although this is not explicitly defined in the text or formula. For this loss to meaningfully capture the “continuous trajectory of neurodegeneration”, it is mathematically critical that the samples within each minibatch exhibit high variance in their ground-truth labels ($y$). If a batch of samples is randomly composed of samples with similar MoCA scores, the correlation calculation will be unstable or uninformative. The authors need to clarify if a specific data sampling strategy (e. g. , stratified sampling across disease severity) was employed to ensure sufficient label variance within each minibatch.
4.Hyperparameter Complexity and Training Stability The framework relies heavily on numerous hyperparameters to balance representation disentanglement, knowledge distillation, and semantic learning. Integrating so many distinct loss functions introduces significant optimization difficulty and raises concerns about the robustness of the model. The authors should provide a comprehensive ablation study or hyperparameter sensitivity analysis to justify how the trade-offs between these loss weights were balanced and to demonstrate that the model is not overly sensitive to these specific empirical settings.
5.Justification for Biological Validity (Section 3.4) In Section 3.4, the methodology for evaluating the “biological validity” of the extracted features by predicting structural volume is confusing. The authors do not provide sufficient justification for why and how predicting structural volume directly equates to biological validity in this specific context. Furthermore, there are no references provided to demonstrate whether this evaluation method has been established or validated in prior literature. Further explanation and appropriate citations are required.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the proposed EyeDEM framework demonstrates strong clinical motivation and impressive standalone classification performance through its innovative latent disentanglement, potential methodological flaws and a lack of technical clarity undermine the paper’s rigor.
- 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 author have provide detailed response to initial comments during the rebuttal.
Review #2
- Please describe the contribution of the paper
1.This paper proposes a cross-modal alignment and distillation framework EyeDEM that transfers structural brain priors into a retinal-only screening model for dementia.
2.The idea of leveraging paired eye-brain data to bridge peripheral biomarkers and central pathology is conceptually interesting and clinically meaningful.
3.The framework is technically well-structured, combining disentanglement, multi-level alignment, and knowledge distillation. The experimental results demonstrate competitive performance.
- 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 addresses an important and clinically relevant problem that using retinal imaging for non-invasive dementia screening. The idea of leveraging eye-brain biological coupling is novel and well motivated. The proposed framework is technically comprehensive, the teacher-student paradigm that train with brain and infer with eye is well-suited for scalable dementia screening.
2.The three-stage pipeline (disentanglement to multi-level alignment to distillation) forms a logical progression. Each component addresses a specific problem.
3.The experimental results are consistently outperform several competitive baselines across both internal and external datasets.
4.The paper validate interpretability of EyeDEM by using learned retinal features to predict structural volumes of seven dementia-associated brain ROIs without additional training. The left hippocampus correlation is biologically plausible and provides meaningful evidence that the model internalizes genuine neurodegenerative signals.
- 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 authors state that MRI is “fundamentally unviable for large-scale population screening due to high costs and invasive nature.” This contains a factual error: MRI is a non-invasive modality by definition. The characterization of invasiveness applies to CSF analysis or PET, but not to MRI. More critically, this claim is logically inconsistent with the authors’ own methodology because the EyeDEM also relies on T1-weighted MRI for BrainPipe construction and distillation training. The authors appear to intend that MRI is unavailable at inference time, but this distinction is never explicitly drawn. The Introduction need to be revised.
2.The entire joint optimization, including disentanglement, MLA, and DCKD, is trained on only 137 paired subjects. With 5-fold cross-validation, each training fold contains approximately 110 samples. Supporting three alignment objectives (Eq. 1–5) and three distillation objectives (Eq. 6) simultaneously on this sample size raises serious concerns about statistical reliability and overfitting.
3.The framework relies on several empirically chosen hyperparameters in the DCKD loss, yet the paper does not provide any sensitivity analysis or justification for these choices. Given that the overall objective is a weighted combination of multiple competing terms, the final performance may be highly dependent on the specific parameter configuration. I’d suggest the paper report ablation results over different weight settings or sensitivity curves.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The paper addresses an important and clinically relevant problem. The idea of leveraging eye-brain biological coupling is novel and well-motivated. The proposed framework is technically comprehensive, integrating disentanglement, multi-level alignment, and distillation in a unified pipeline. The experimental results are strong and consistently outperform several competitive baselines across both internal and external datasets. The inclusion of ablation studies and visualization further supports the design choices and provides some level of interpretability.
2.The paper describes MRI and PET as having an “invasive nature” and being “fundamentally unviable for large-scale population screening,” which is not entirely accurate and may be misleading because MRI is a non-invasive imaging modality. Moreover, the argument presents a slight inconsistency, as MRI is also used within the proposed framework as a key modality for training and knowledge transfer. The Introduction should be revised. An ideal introduction should accurately distinguish between the training-time use of MRI as a privileged modality and the inference-time constraint of retinal-only input, rather than making an overgeneralized claim about MRI’s viability.
3.The framework relies on several empirically chosen hyperparameters; it would benefit from additional sensitivity analysis to better understand the robustness of the model under different parameter settings.
- 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’ rebuttal addresses my major concerns.
Review #3
- Please describe the contribution of the paper
This paper proposes EyeDEM, a cross-modal training framework for dementia screening from retinal OCTA. The method uses paired retinal and brain MRI data during training to align eye and brain representations and distill brain structural priors into an eye-only screening model. At inference time, the model requires retinal OCTA alone. The evaluation includes internal cross-validation, an external retinal-only test set, ablations, and an analysis relating learned retinal representations to MRI-derived regional brain volumes.
- 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 paper addresses a clinically meaningful problem. Retinal imaging is considerably more practical than MRI, PET, or CSF for screening, so using paired eye-brain data during training to improve retinal-only dementia prediction is a relevant and translationally interesting direction. I also think the paper is strongest when viewed through this application lens. The overall pipeline is coherent. The disentanglement, multi-level alignment, and distillation components are well aligned with the goal of transferring brain structural priors into a standalone retinal model. I also appreciate that the evaluation goes beyond classification metrics alone: the external test set is important for this kind of screening paper, and the ROI-level correlation analysis is a useful attempt to support biological plausibility. The paper is also reasonably complete experimentally. It compares against both generic image backbones and OCTA-specific dementia baselines, and the ablation study helps show that the different components contribute to the final result.
- 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 clinical setup is not fully specified. In particular, the cohort composition and label definition across different stages of training and evaluation remain somewhat unclear, especially given the use of multiple dementia-related cohorts and the paired internal cohort with CSF-derived ground truth. 2.The paired cohort is relatively small for the main cross-modal claim, and the split policy is not described clearly enough. It would help to clarify whether evaluation was fully subject-wise and how repeated scans or paired samples were handled across folds. 3.Robustness considerations are under-discussed. OCTA-based biomarkers can be sensitive to image quality, artifacts, and center effects, but the current manuscript provides limited detail on quality control and related confounding factors. 4.Statistical reporting is somewhat limited. The paper reports useful summary metrics, but uncertainty, calibration, confidence intervals, and cross-cohort stability are not analyzed in sufficient depth. 5.The interpretability analysis is interesting, but it is better viewed as supportive evidence rather than definitive proof of robust retinal encoding of cerebral degeneration.
- 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 am overall positive on this paper. The clinical motivation is strong, the proposed framework is coherent, and the evaluation includes several elements that matter for this type of work, especially external testing, baseline comparison, and ablation. I also think the idea of using paired eye-brain data during training to improve a standalone retinal screening model is both practically relevant and timely. My reservations are mainly about rigor and reporting rather than about the overall direction. In particular, I would have liked clearer cohort/label definitions, more discussion of OCTA quality and potential confounding, and somewhat stronger statistical reporting. Even with these limitations, I view the paper as slightly above the acceptance threshold for an application-driven submission.
- 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
1.Gradient flow & Mathematical Definitions(R1) During cross-modal interaction (alignment/distillation), BrainPipe is a frozen teacher: its parameters are non-trainable and excluded from the optimizer, and teacher features are detached from the computation graph via stop-gradient, like sg(M_B). Thus, gradients from L_ins flow only to EyePipe and student modules, not vice-versa; “joint optimized” excludes BrainPipe. Regarding notation: tilde(C_B^+) denotes batch-permuted C_B^+, pairing each C_E^+ with another subject’s C_B^+ to estimate the JSD marginal term. H_E and H_{B+E} are the penultimate features from the eye student (f_E) and the multimodal teacher (f_{fus}); L_weak and L_KL follow the definitions in Ref. [9]. 2.Hyperparameters and MoCA Loss (R1 & 2) To ensure stability and prevent overfitting, we pretrain encoders on large-scale datasets (1,046 OCTA; 1,278 MRI) before fine-tuning on the 137 paired subjects. The hyperparameters were selected via grid search, and sensitivity analysis shows robust performance, with AUC variations limited to 0.01–0.015.L_MoCA is per-batch (B=32). As MoCA is continuous, we did not stratify by disease severity, but used class-balanced WeightedRandomSampler per epoch. Within-batch MoCA SD was 4.3–10.0, avoiding near-zero variance and stabilizing Pearson supervision. Ablation results confirm that each loss component is essential, as removing any single module leads to a performance drop (up to 0.043 AUC). 3.Biological Validity (R1): Predicting structural volumes evaluates whether EyeDEM’s eye representations encode information associated with central macro-structural pathology, rather than only superficial eye patterns.MRI is used as privileged information during training to guide the OCTA encoder toward eye-brain biological coupling. Since brain atrophy, especially hippocampal volume loss, is a well-established dementia phenotype, predicting brain volumes from EyeDEM representations without further fine-tuning suggests that the latent space captures shared pathology-related information along the eye-brain axis. This provides an indirect biological explanation for the OCTA encoder’s gain. We will clarify that this is a correlation-based validity analysis, not causal biomarker validation. 4.MRI clarification（R2） We acknowledge the wording error. MRI is non-invasive but costly/low-throughput for screening; PET requires tracer administration and CSF is invasive. MRI is used only as a privileged training modality to build BrainPipe and guide distillation. At inference, EyeDEM uses OCTA alone. 5.Limited data & Reliability （R2 & 3） EyeDEM does not learn encoders solely from 137 paired subjects; larger unpaired cohorts (1046 OCTA, 1278 MRI) are used for independent modality pre-training. In the paired-set stage, optimization mainly updates student-side alignment/distillation modules to internalize eye-brain biological coupling, rather than learning basic image features from scratch. We also conducted subject-level paired bootstrap testing against the strongest baseline (p<0.04, max 95% CI width <0.31,eg., ΔAUC 95% CI [0.008, 0.201]), but omitted it due to space limits, and external OCTA-only validation (AUC 0.761), suggesting gains are not from internal-cohort overfitting. 6.Cohort Details (R3) Labels were defined as dementia vs. NC. AD follows NIA-AA with amyloid positivity; VAD follows VASCOG with MRI-supported cerebrovascular disease; NC excludes dementia, major neurological disease, and retinal/optic nerve disorders. To ensure statistical rigor, all dataset splits were strictly subject-wise, with bilateral OCTA scans or paired samples from the same participant maintained within the same fold to prevent data leakage. We did not utilize repeated scans for training or evaluation. Instead, we implemented a strict selection policy to retain only the single highest-quality OCTA image for each subject, defined as a software-derived image quality score > 6/10.We will include cohort description.
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 paper proposes a cross-modal framework for dementia screening using retinal OCTA with paired eye-brain data during training. All reviewers agree on the clinical relevance and coherent pipeline design. R1 and R2 raise significant concerns about gradient flow ambiguity, incomplete mathematical definitions, hyperparameter complexity, and small sample size (137 paired subjects). R3 is more positive, appreciating the external testing, baseline comparison, and ablation study, but notes limitations in cohort specification, statistical reporting, and robustness considerations.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal addresses the main technical and conceptual concerns and clarifies that MRI is used only as privileged information during training while inference relies solely on OCTA, which aligns with the stated screening motivation. The overall contribution remains novel and clinically relevant, and the additional clarifications increase confidence in the robustness and validity of the reported gains
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.
The authors have satisfactorily addressed the Reviewers’ comments and the proposed methodology sounds promising.
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.
The authors have provided detailed responses to all reviewers’ main concerns during the rebuttal.