List of Papers Browse by Subject Areas Author List
Abstract
Mobile screening is increasingly needed for frequent, low-burden assessment of cognitive impairment in communities and households, and eye tracking offers scalable oculomotor signatures for this purpose. However, mobile gaze estimation is fragile under coupled shifts from aging-related appearance changes, uncontrolled capture noise, and fluctuating engagement, which can corrupt downstream oculomotor features. We propose CoGaze, a closed-loop gaze reasoning framework for mobile videos of older adults. CoGaze formulates gaze estimation as structured alignment between ocular appearance and facial geometric context. It stabilizes ocular representations with a prototype–anchor dictionary retrieved by normalized cosine attention, refines eye–face evidence via reciprocal cross-attention, and enforces temporal alignment using differential spatiotemporal attention with a segment-level consistency objective to reduce jitter and drift without blurring true saccades. We collect a deployment-oriented multi-center dataset from 7 sites (July–December 2025) with 119 gaze-task subjects and 114 clinically labeled assessments. CoGaze achieves mean error 1.29 cm/2.03 cm on phone/tablet and improves screening to 85.1% accuracy (5-fold CV). The code is available at: https://github.com/xiangxuebei/CoGaze
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3065_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xiangxuebei/CoGaze
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YuJia_CoGaze_MICCAI2026,
author = { Yu, Jiahui AND Du, Tao AND Wang, Jinfu AND Wu, You AND Lou, Xiaowen AND Pang, Ting AND Xu, Xin},
title = { { CoGaze: Closed-Loop Eye-Face Alignment for Multi-center Mobile Gaze-Based Cognitive Screening in Older Adults } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes CoGaze, a closed-loop eye–face alignment framework for robust mobile gaze estimation in older adults. The method integrates an ocular prototype–anchor dictionary, reciprocal cross-attention between eye and facial features, and differential spatiotemporal attention to reduce jitter and drift in gaze trajectories. The authors also introduce a multi-center mobile dataset and demonstrate that improved gaze estimation leads to modest gains in downstream cognitive screening performance based on derived oculomotor features.
- 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 presents a technically strong and well-motivated contribution to mobile gaze estimation under challenging real-world conditions. The proposed closed-loop eye–face alignment mechanism is novel and addresses practical issues such as aging-related appearance variation, motion noise, and temporal instability. The method demonstrates consistent improvements over strong baselines across multiple devices and tasks, with clear gains in both localization error and robustness metrics. The collection of a multi-center dataset is also a valuable contribution, and the focus on deployment-oriented settings (mobile devices, older populations) is timely and relevant. Overall, the work is well-executed from a machine learning and systems perspective.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The main limitation of the paper is that the clinical contribution is relatively limited compared to the methodological focus. While the work is framed as cognitive screening, the primary advances lie in gaze estimation, and the downstream clinical analysis remains shallow. The paper does not provide detailed insight into which oculomotor features are most informative, how they relate to cognitive domains, or how the model could support clinical decision-making.
The evaluation of cognitive screening performance is also limited. The reported accuracy (85.1%) is obtained via 5-fold cross-validation on a relatively small dataset (114 subjects), without external validation. This raises concerns about generalizability, especially given class imbalance and potential site-specific biases. Furthermore, the improvement in screening performance appears modest relative to the substantial gains in gaze estimation, and it is not clearly demonstrated whether these gains translate into meaningful clinical benefit.
In addition, the paper lacks a discussion of limitations. Important aspects such as dataset size, cohort composition, potential biases, and deployment constraints are not critically examined. A more balanced discussion would strengthen the paper.
Finally, while the method is technically sophisticated, it remains unclear how much of the downstream improvement is attributable to the proposed alignment mechanism versus general improvements in gaze accuracy or feature extraction.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 paper presents a strong and well-executed methodological contribution to mobile gaze estimation, with clear improvements over existing approaches. However, the clinical relevance is limited, and the cognitive screening component is not sufficiently developed or validated to support strong translational claims. In particular, the lack of external validation, limited analysis of clinical features, and absence of a discussion of limitations weaken the overall impact. While the work is promising, it would benefit from stronger clinical grounding and more rigorous evaluation of the screening task.
- 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 rebuttal usefully clarifies several methodological points, especially that downstream classification used subject-level stratified 5-fold cross-validation and that all compared methods shared the same folds, features, and classifier. This addresses possible concerns about within-subject leakage and helps explain the source of the reported downstream gains.
However, my main concern regarding clinical validation and generalizability remains only partially addressed. The current submission still lacks external validation, held-out-site/device evaluation, or leave-one-center-out analysis, despite being framed as a multi-center, deployment-oriented study. The cognitive-screening results are based on internal cross-validation on a relatively small cohort of 114 subjects, with limited dementia cases, which weakens the strength of the translational claims.
Overall, I still view the paper as a technically strong and promising methodological contribution with practical relevance, but not one that is sufficiently validated clinically to merit stronger enthusiasm. I therefore remain on the accept side, but only as a weak accept.
Review #2
- Please describe the contribution of the paper
This paper addresses a meaningful and timely problem by connecting mobile gaze estimation with downstream cognitive screening in older adults, and the multi-center dataset is a notable strength. However, while the results are promising, the current manuscript does not yet provide enough clarity and depth in the evaluation protocol, downstream classification setting, and ablation study to fully support the strength of its claims.
- 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 studies a meaningful problem and has clear practical relevance, especially for mobile cognitive screening in older adults. A clear strength is the multi-center dataset collected from seven sites, which makes the work more valuable than papers evaluated only on small or single-site data. I also think it is a plus that the paper connects gaze estimation with downstream cognitive screening, rather than stopping at the intermediate task. Overall, the motivation is strong, the application setting is realistic, and the reported results are promising.
- 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 paper is framed as a multi-center, deployment-oriented study, but the evaluation setup is still not clear enough. I was not fully sure how the subject splits were done, whether the cross-validation was strictly subject-independent, or how much the results reflect true cross-center robustness. For this kind of application, those details matter a lot. 2.The cognitive-screening part also needs more explanation. The paper shows that screening accuracy improves when using gaze features from CoGaze, but the downstream pipeline is only described at a fairly high level. It is therefore hard to tell whether the gain really comes from better gaze estimation, from the classifier itself, or from the particular split of a relatively small and imbalanced cohort. 3.The ablation study feels too thin for a method with this many components. The model includes the ocular prior dictionary, the multi-scale face branch, the eye-face interaction module, and the temporal alignment design, but only part of these are actually examined in the ablation table. Because of that, it is still not very clear which pieces are doing the heavy lifting. 4.I also felt that some of the presentation is a bit stronger than the evidence currently supports. The overall idea is interesting, but terms like “closed-loop reasoning” and the broader claims about robust mobile screening would be more convincing with stronger generalization tests and more detailed analysis across centers, devices, or subject groups.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
This paper addresses a meaningful and timely problem by connecting mobile gaze estimation with downstream cognitive screening in older adults, and the multi-center dataset is a notable strength.
- 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 recommend acceptance, as the rebuttal sufficiently clarified the main concerns regarding the evaluation protocol and downstream screening setup, although some limitations remain in the depth of robustness analysis and the ablation study.
Review #3
- Please describe the contribution of the paper
The authors propose a closed-loop eye-face alignment framework for gaze inference tailored to mobile cognitive screening in older adults, and construct a mobile gaze prediction dataset comprising 119 participants.
- 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 study constructs a relatively comprehensive dataset involving seven scenarios, dual mobile devices (smartphones and tablets), gaze data from 119 participants, and clinical cognitive labels from 114 participants, which strengthens the empirical basis of the work. (2) The task focuses on older adults, which distinguishes it from previous gaze estimation models developed for more general populations or 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.
(1) The method mainly relies on a basic TSM for temporal modeling, without exploring stronger video modeling architectures such as Transformers. Moreover, since the main difference from conventional gaze estimation methods lies in the face-eye cross-attention design, the manuscript should better explain why this design is sufficient to handle pose jitter, illumination changes, occlusion, age-related appearance variations, and attention fluctuations.
(2) The experimental section lacks comparisons with several classical gaze estimation methods.
(3) It would be helpful to evaluate the model on standard datasets such as GazeCapture to further verify its generalizability. In addition, since the method targets older adults, the manuscript should more clearly define the age range of the intended population.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
This assessment is mainly based on prior gaze estimation studies, the method proposed in this work, its experimental results, and the released code.
- 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
We sincerely thank all reviewers for their constructive feedback and for recognizing the methodological innovation and practical significance of our work (R1–R3), the value of the multi-center dataset and application scenario (R1–R3), and the overall execution of the study (R1–R2). Below, we respond to the reviewers’ comments. 1) Clinical scope [R1.1, R1.3, R2.4]. CoGaze is not intended to replace clinical diagnosis or standard neuropsychological assessment. Instead, it is a lightweight mobile gaze-measurement tool for early cognitive screening, using gaze-derived features as objective and quantitative indicators to detect subtle abnormalities that may be missed by conventional scales. We extracted interpretable oculomotor descriptors, including fixation, saccade, and latency features, which have been linked in prior studies to attention, visual search, and executive function in early cognitive impairment. CoGaze is designed for low-burden screening in outpatient clinics, communities, and nursing homes, using consumer-grade mobile devices rather than specialized eye trackers. In our cohort, 61.4% of participants reported subjective cognitive complaints; together with MCI participants, they represent a key target population for early screening. We will also discuss limitations rigorously and conduct larger external multi-center validation, cross-device testing, and subgroup analyses in future work. 2) Evaluation validity [R1.2, R1.4, R2.1, R2.2]. We will clarify that gaze estimation and cognitive screening were conducted in two separate stages. After training each gaze estimator on the gaze-task data, we fixed it to extract gaze trajectories from the cognitive-screening videos and converted them into subject-level oculomotor features. Downstream classification adopted a strict subject-level stratified 5-fold cross-validation protocol according to the NCI/MCI/dementia distribution, ensuring that all data from the same participant remained within the same fold. Importantly, all compared methods used the same folds, feature categories, and classifier; therefore, the downstream improvement mainly reflects the more stable gaze-derived phenotypes produced by CoGaze, rather than effects from the classifier or class distribution. In addition, we performed t-tests on the classification results, showing that our method significantly outperformed the other methods (p < 0.05). 3) Robustness Mechanism [R2.3, R3.1]. In older-adult mobile acquisition, eye crops alone are unreliable: pose changes geometry; illumination/glasses alter texture; occlusion/blinking corrupts pupil/eyelid cues; aging enlarges appearance variation; attention lapses cause instability. CoGaze treats eye-face interaction as calibration, not fusion. Ocular prior maps eye evidence to normalized anchors; multi-scale face tokens provide head-pose and geometry; EFCA reweights ocular evidence and reduces reliance on invalid tokens. ADM and segment consistency suppress jitter/drift while preserving true saccades, and TSM targets short-term stability. EFCA is the core interaction form, whose effect depends on reliable ocular and facial evidence. Thus, ablation keeps EFCA and removes Pd, Msc, or ADM to test what makes the interaction reliable. The full model improves Pix200 from 0.815 to 0.874 and reduces fixation error from 1.52 to 1.32; Pd and Msc gains show anchors and facial context are complementary. 4) Baseline Comparison [R3.2, R3.3]:We compared CoGaze with iTracker, BlazeGaze, ST-WSGE, and FIFA, covering 2D/3D-based and temporal sequence methods. On GazeCapture, CoGaze also achieved the best performance, with an average error of 1.65 cm across phones and tablets, reducing the error by 0.29 cm compared with the classical gaze-estimation method AFF-Net. These results were omitted due to manuscript space limits. In addition, our target population was aged 50–90 years (mean: 72.7 ± 6.7), which will be clarified in the manuscript.
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.
This submission proposes CoGaze, a closed-loop gaze reasoning framework for mobile cognitive screening and collected a multi-center dataset of older adults. Two reviewers (R1, R2) vote for weak accept and one reviewer (R3) votes for weak reject. Concerns remain regarding clinical validation, evaluation clarity, and completeness of experimental comparisons. The key points to address in the rebuttal are as follows: R1 raises a major concern that the clinical relevance of the work is not sufficiently substantiated. R2 points out the evaluation setup lacks clarity and the ablation study is too limited given the number of componentsR3 highlights the lack of comparisons with classical gaze estimation methods and evaluation on standard datasets such as, GazeCapture.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The authors provided additional clarifications, yet there are still some concerns about clinical relevance, robustness / generalizability analysis and the ablation study such as evaluation on public datasets.
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 main limitations concern the strength of the clinical and translational evidence. Although the paper is framed around cognitive screening, the strongest contribution is methodological and lies primarily in gaze estimation. The downstream screening analysis remains relatively shallow: the cohort is modest in size, the dementia class is small, and the reported screening performance is based on internal 5-fold cross-validation rather than external validation, held-out-site testing, held-out-device testing, or leave-one-center-out evaluation. As a result, the evidence does not yet fully support strong claims about deployment robustness or clinical generalizability. The ablation study also remains somewhat limited for a method with several interacting components, since not all elements are independently isolated. The rebuttal improved the assessment by clarifying important points about the evaluation protocol. In particular, it clarified that downstream classification used strict subject-level stratified 5-fold cross-validation, and that all compared methods used the same folds, feature categories, and classifier. This addresses concerns about possible within-subject leakage and makes the downstream comparison more credible. The rebuttal also clarified the intended clinical scope as early mobile screening rather than replacement of clinical diagnosis, and explained the roles of the ocular prior, multi-scale face representation, EFCA, and ADM. However, the rebuttal only partly addresses the concerns about clinical validation and cross-center/device robustness. The additional standard-dataset result mentioned in the rebuttal is informative but was not treated as decisive for the decision, since it was not part of the submitted manuscript. The reviewer opinions were mixed but leaned positive after rebuttal. Two reviewers initially viewed the paper as a weak accept and maintained or moved to acceptance after the rebuttal, while one reviewer raised a weak-reject position focused mainly on missing comparisons and generalization evidence. I consider the latter concerns valid, especially for the translational framing, but not sufficient to outweigh the methodological contribution, the real-world older-adult dataset, and the consistent experimental improvements.
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 submitted manuscript presents a technically solid mobile gaze-estimation framework with a valuable multi-center older-adult dataset and consistent improvements in gaze accuracy and downstream screening. However, the clinical validation remains limited, especially without external or leave-one-center-out evaluation, so the translational claims should be moderated.
