List of Papers Browse by Subject Areas Author List
Abstract
Rapid stroke screening is crucial in remote and resource-limited settings where timely medical expertise may be unavailable. Facial palsy, a visible and early indicator of stroke, offers a practical and non-invasive basis for assessment, motivating the need for automated and reliable recognition methods. To address this need, we propose DP-Face, a video-based deep learning framework that captures dynamic facial asymmetry while preserving patient privacy. Rather than relying on raw RGB appearance, DP-Face integrates structural cues derived from facial landmarks and wrinkle patterns with temporal muscle dynamics modeled through facial Action Units. To promote reproducibility, we also construct Palsy-330, the largest publicly available video dataset to date for facial palsy classification, comprising 330 subjects. Evaluated under a patient-independent protocol, DP-Face outperforms existing baselines, demonstrating the promise of dynamic, privacy-preserving facial analysis for rapid stroke screening. The source code and dataset are available at
https://github.com/ChangYuance/DP-Face.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2957_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ChangYuance/DP-Face
Link to the Dataset(s)
Palsy-330 dataset: https://pan.baidu.com/s/1tUpvrToivws9mSsGXN-S4w?pwd=cwc7
BibTex
@InProceedings{ChaYua_PrivacyPreserving_MICCAI2026,
author = { Chang, Yuance AND Liu, Bang AND Ding, Han AND Zhao, Cui AND Wang, Fei AND Wang, Ge AND Xi, Wei},
title = { { Privacy-Preserving Video-Based Facial Palsy Recognition via Dynamic Facial Action Modeling } },
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 a video-based deep learning framework, DP-Face, for the recognition of facial palsy. The method generates privacy-preserving images by combining facial keypoint connections and wrinkle features, while leveraging facial action units to capture dynamic sequences of muscle movements. In the feature fusion stage, the model adopts a FiLM-based modulation mechanism to integrate AU dynamic features into appearance features. Furthermore, this paper constructs Palsy-330, currently the largest video-level facial palsy classification dataset, which includes 330 subjects.
- 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.Introduces Palsy-330, the largest video-level dataset (330 subjects) to date for facial palsy classification. 2.The proposed PP images effectively eliminate identity-discriminative information. 3.Utilizes the FiLM mechanism for multimodal modulation, which is empirically proven by ablation experiments to be superior to standard concatenation and attention strategies.
- 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.Lack of innovation: The DP-Face framework is essentially an assembly of standard visual and sequential modules. 2.Typographical error in Table 3: In Table 3, the Specificity (Spec. ) value 52.5 of for DP-Face is incorrectly bolded. The highest value in this column is 70.0, achieved by the FP-VGGFace.
- 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?
Due to a lack of innovation and rigor in data presentation, I recommend a Weak Reject.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors’ rebuttal can partially solve my concerns. While I understand the clinical trade-off of prioritizing high recall over specificity for early stroke screening, a specificity of 52.5% still poses a risk of introducing too many false positives in real-world clinical workflows. However, the application-driven motivation is clear, and the integration of privacy-preserving features for screening shows good clinical novelty and practical value. The evaluation on a real multi-center dataset is also appreciated. In summary, considering the overall contribution, I decide to update my rating to Weak Accept.
Review #2
- Please describe the contribution of the paper
This paper presents DP-Face, a video-based framework for facial palsy recognition aimed at early stroke screening. The method combines privacy-preserving facial representations derived from landmarks and wrinkle patterns with temporal modeling of facial dynamics using Action Units (AUs). A multimodal fusion strategy based on FiLM integrates appearance and motion cues. Additionally, the authors introduce Palsy-330, a new dataset of 330 participants collected across multiple hospitals. Experimental results show moderate improvements over existing baselines and some generalization capability on an external dataset.
- 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 relevant and impactful problem: early stroke screening through facial palsy detection in resource-limited settings.
-) The introduction of the Palsy-330 dataset is a valuable contribution, as it is relatively large for this task and collected in real clinical environments.
-) The use of participant-independent cross-validation is appropriate and avoids identity leakage during evaluation.
-) The proposed framework combines structural and temporal information in a reasonable way, leveraging both geometric representations and Action Units.
-) The idea of privacy-preserving representations is relevant, especially in medical applications involving sensitive patient 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.
-) The dataset description lacks important details, such as the distribution of samples per participant and more comprehensive statistics, making it difficult to assess potential biases.
-) The decision on using [27] for facial features detection is not justified. Why are you selecting this method and not others? And do you have any clous on how it can affect the overall pipeline?
-) The annotation protocol is insufficiently described, with no information about the number of clinicians, inter-rater agreement, or precise criteria for labeling facial palsy.
-) The temporal modeling design is not well justified, particularly the choice of sampling only 16 frames from sequences that are significantly longer, and it is unclear how temporal coverage is ensured. If you change this parameter, how does it affect the performance? And is there overlap between consecutive time windows?
-) The privacy-preserving claim is weakly supported, as identity recognition accuracy remains relatively high and no formal privacy evaluation or comparison to standard anonymization techniques is provided.
-) The performance improvements over prior methods are relatively modest and lack statistical significance analysis.
-) The evaluation does not fully address clinical relevance, as it focuses only on binary classification without considering severity levels or comparison with expert assessments.
-) Some implementation and preprocessing details are unclear, including specifics of AU extraction reliability and aspects of the training pipeline.
- 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?
Overall, the paper addresses a clinically relevant problem and introduces a potentially valuable dataset together with a privacy-aware video-based framework. The integration of structural and temporal features is reasonable and supported by experimental results. However, several important aspects limit the strength of the contribution, including insufficient details about the dataset and annotation protocol, limited validation of the privacy claims, and relatively modest performance gains over existing methods. Additionally, the methodological novelty is incremental. Addressing these issues, particularly improving the clarity and rigor of the dataset description and providing stronger validation, would significantly strengthen the work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 rebuttal addresses my concerns by clarifying the dataset structure and emphasizing the reproducibility effort; however, the methodological contribution remains relatively limited and the lack of stronger baselines and annotation reliability analysis still weakens the overall impact.
Review #3
- Please describe the contribution of the paper
This paper presents DP-Face, a video-based framework for facial palsy recognition aimed at enabling rapid stroke screening.
- 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.
DP-Face combines a) a privacy-preserving structural representation (facial landmarks + wrinkle maps) with b) temporal facial Action Unit (AU) modeling, followed by multimodal fusion using FiLM. In addition, the authors introduce Palsy-330, a 330-subject dataset collected under clinical conditions with a standardized smile task. Experimental results show moderate improvements over prior image- and video-based baselines under a subject-independent protocol
- 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.Incremental novelty in modeling despite reasonable integration. The core modeling components—AU-based representation, wrinkle/nasolabial modeling, and temporal dynamics—have all been explored in prior work. For example, AU-based facial palsy analysis and wrinkle-based asymmetry quantification have been previously studied, and recent works already combine structured facial representations with learned features. As such, the main contribution lies in the integration of privacy-preserving structural features with temporal AU modeling, rather than a fundamentally new modeling paradigm. This is still a reasonable contribution for MICCAI given the application context, but the paper would benefit from more precise positioning relative to prior work and a clearer articulation of what is novel.
2.The proposed “privacy-preserving” representation reduces identity classification accuracy from 99% to 57.6%, but this is still far above chance, suggesting residual identity leakage. There is no formal privacy framework (e.g., differential privacy, re-identification risk) used, and no comparison to alternative anonymization techniques (e.g., face de-identification, GAN-based anonymization). Some recent work (e.g., https://doi.org/10.1007/978-3-031-96625-5_26) also performs explicit de-identification via motion transfer to synthetic identities and evaluates privacy–utility tradeoffs more rigorously; the current evaluation is relatively limited.
3.Clinical framing (stroke screening) is somewhat overstated. The paper is motivated by stroke screening, but the task evaluated is facial palsy classification (present vs. absent). While facial palsy is an important indicator, real-world stroke triage involves multiple symptoms and differential diagnoses (e.g., Bell’s palsy). The absence of evaluation in a true stroke or triage setting limits the strength of the clinical claims. This does not invalidate the work, but the paper would benefit from more cautious positioning, framing the contribution as facial palsy detection toward stroke screening rather than a direct solution.
- 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?
This paper provides a useful and well-executed contribution to video-based facial palsy analysis, particularly through the release of a relatively large public dataset and a practical privacy-aware modeling pipeline. While the methodological novelty is incremental and the privacy/clinical claims could be better substantiated, the work is technically sound, reproducible, and relevant to the MICCAI community. With improved positioning and discussion of related work, this paper would make a valuable, application-oriented contribution.
- 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.
Some, but not all, of my concerns were addressed. Still leaning slightly towards the accept side, considering the privacy-preserving domain is relatively underexplored for neurological diagnoses
Author Feedback
Q1: Methodological Innovation and Contribution (R1, R3, R4) A: Our contribution is application-driven (MICCAI-oriented) rather than proposing a new backbone. The novelty lies in integrating existing components to address a real clinical problem: privacy-preserving video analysis for fast, early stroke screening. We are the first to combine PP images (de-identified geometry) with AU temporal dynamics for palsy detection. PP images remove identity cues (see Q3). As Table 2 shows, DP-Face achieves 75.1% accuracy on real data from ten Chinese hospitals, demonstrating strong potential for both home and clinical screening, even when compared to methods that directly use raw RGB images.
Q2: Dataset Details (R3) A: Sampling: Each subject recorded one 4s smiling video (20 fps, ~80 frames). We apply 5 temporal offsets for uniform sampling, generating 4~5 samples per subject (mean±std in Table 1).
Annotation: Primary labels were provided by the collecting physicians, verified locally before uploading. The entire dataset was reviewed by a Chief Physician (15+ years experience) for consistency.
Other Details: Data were collected from 10 hospitals across central and western China, each contributing 2–101 subjects (mean 33.4, std 30.7). The top 3 hospitals contribute 65.9%, the last 4 less than 6%, reflecting real-world imbalance and providing a natural non-IID setting for further research topics such as domain adaptation and federated learning. The above metadata, along with additional information (e.g., age, gender), will be released together with the dataset to support the community.
Q3: Privacy-Preserving Validation (R3, R4) A: We update the privacy-preserving validation experiment. We instead report ReID performance to provide stronger evidence. We extract features from RGB and PP images using separate ResNet-50 networks and train them with a triplet loss. For comparison, we also run the same experiment with RGB images only. Rank-1 accuracy drops from 100% (RGB-only) to 5.11% (PP→RGB, 100-person gallery, chance 1%), showing PP features prevent re-identification while retaining diagnostic utility. This also indicates the previous 57.6% classification result reflects memorization of “line structures” rather than identity.
Q4: Stroke Screening vs. Palsy Detection (R3, R4) A: Facial palsy is a key BEFAST indicator for early stroke screening, but our model detects symptoms only. A positive result indicates the need for further medical consultation, not a definitive stroke diagnosis. We agree and will revise the description as “Facial Palsy Detection towards Stroke Screening”.
Q5: Choice of STAR (R3) A: Accurate facial landmark detection is critical for reliable structural cues in facial palsy analysis. MediaPipe [2] is widely used but often produces overly symmetric 3D landmarks under severe asymmetry, reducing reliability. In contrast, STAR employs Self-adapTive Ambiguity Reduction (STAR) loss to reduce annotation ambiguity and produce more robust landmark features. It runs >20 fps on GPU, enabling real-time clinical application.
Q6: Spec. and Table 3 Error (R1) A: We apologize for the typo error and will correct it in Table 3.While FP-VGGFace, which uses raw RGB images, has the highest specificity (70.0%), our DP-Face achieves lower specificity (52.5%) but the highest Recall (94.5%) and Accuracy (87.3%). We augur that due to the severe class imbalance in MEEI (Palsy:Negative=5:1), misclassifying a few healthy subjects can significantly reduce specificity. For stroke screening, we think high recall—minimizing missed diagnoses—is more important over specificity.
Q7: AU Extraction Details and Sampling Rate (R3) A: We use Py-Feat—a popular Python toolbox pre-trained on large FACS datasets—to extract AU features offline from RGB frames, producing 20-dim intensity vectors per frame as model inputs. And we sample 16 frames to balance efficiency and performance. Table 4 (ablation) shows reducing to 8 or 4 frames decreases accuracy.
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 addresses a clinically relevant problem in video-based facial palsy recognition and introduces Palsy-330, a valuable multi-subject dataset collected under clinical conditions. Reviewers appreciated the practical motivation, dataset contribution, subject-independent evaluation protocol, privacy-aware representation, and integration of structural facial cues with temporal AU modeling. However, several concerns should be addressed in rebuttal. The reviewers consistently noted that the methodological novelty is incremental, as many components are based on existing facial representation and temporal modeling ideas. They also raised concerns about incomplete dataset and annotation details, limited validation of the privacy-preserving claim given residual identity leakage, modest performance gains, and somewhat overstated clinical framing as stroke screening rather than facial palsy detection. A focused rebuttal clarifying these and other points raised by reviewers would help reassess the paper’s contribution.
- 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.
DP-Face targets a clinically relevant problem — privacy-preserving, video-based facial palsy recognition toward stroke screening — and its main strengths are the application-driven framing and Palsy-330, a large multi-hospital video dataset collected under real clinical conditions, evaluated under a subject-independent protocol. Reviewers consistently valued the dataset and practical motivation, while noting the methodological novelty is incremental (an integration of known components). The rebuttal clarified dataset/annotation details, strengthened the privacy validation, and appropriately tempered the clinical framing. Two reviewers moved to Accept. For camera-ready, authors should incorporate the rebuttal clarifications, soften the clinical claim, correct the Table 3 error, and release the dataset and code as committed.
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.
After the rebuttal, two reviewers recommended acceptance, while one reviewer maintained a Reject recommendation. While the paper has some limitations regarding technical contribution, it also demonstrates several strengths, including a valuable public clinical video dataset and a practical framework design. Therefore, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.
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 rebuttal addressed several concerns in a useful way. It clarified the dataset structure, sampling strategy, hospital distribution, annotation workflow, AU extraction, choice of landmark detector, and the intended clinical framing as “facial palsy detection towards stroke screening” rather than direct stroke screening. It also acknowledged the specificity issue on MEEI and the typo in Table 3.The additional ReID analysis described in the rebuttal helps support the authors’ interpretation that the privacy-preserving representation reduces re-identification risk, although this remains less strong than a formal privacy evaluation and was not part of the original manuscript. The reviewer opinions after rebuttal remain mixed. R1 moved from Weak Reject to Accept, while still noting that low specificity may create many false positives in a practical screening workflow. R4 remained on the accept side, emphasizing the relevance of privacy-preserving neurological video analysis despite only partial resolution of concerns. R3 maintained a Reject recommendation, mainly because the methodological contribution remains limited and because stronger baselines and annotation reliability analysis are still missing. I weigh these points as follows: the unresolved weaknesses are real and should be discussed more carefully in the camera-ready version, but they do not fully outweigh the dataset contribution, the subject-independent evaluation, the practical privacy-aware formulation, and the reasonable empirical evidence. Overall, the paper is recommended for acceptance as a marginal but acceptable MICCAI contribution.
