Abstract

Deep learning for facial-based elderly malnutrition screening shows preliminary feasibility, but clinical data scarcity limits diagnostic performance.

Although diffusion models are widely used for medical image augmentation, their application to facial malnutrition screening is limited by insufficient control over malnutrition-related phenotypes.

To address these limitations, we propose NutriDiff, which fine-tunes Stable Diffusion using view-class LoRA adapters and filters malnutrition-irrelevant synthetic samples through facial morphological feature extraction.

In few-shot testing on 92 elderly subjects, NutriDiff achieves AUC 0.984 and MCC 0.854, outperforming baselines and offering a potential approach for elderly malnutrition screening.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4999_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{YanHan_NutriDiff_MICCAI2026,
        author = { Yang, Hanqiu AND Zhang, Chenming AND Shen, Zhijie AND Wang, Xue AND Bai, Hongtao AND He, Lili},
        title = { { NutriDiff: Phenotype-Consistent Diffusion Augmentation for Malnutrition Screening } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces a diffusion-based data augmentation framework tailored for facial malnutrition screening in elderly patients. The method combines view-conditioned LoRA fine-tuning of Stable Diffusion with a novel post-generation filtering step based on handcrafted facial morphological descriptors (temporal, periorbital, malar, jawline regions) and Mahalanobis distance to remove clinically inconsistent samples. This pipeline is integrated into a downstream classifier (CLIP-based), and the authors demonstrate that phenotype-aware filtering of synthetic data significantly improves generalization in a few-shot clinical setting.

  • 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 subject of elderly malnutrition screening via facial images is clinically relevant and underexplored.
    • The paper introduces a clinically grounded filtering mechanism that enforces phenotype consistency in diffusion-generated samples, addressing label noise which is a known limitation in medical image synthesis.
    • The authors propose a multi-view, view-conditioned LoRA adaptation strategy, enabling fine-grained control of facial pose and nutritional state without full model fine-tuning.
    • The paper combines generative modeling with domain-specific feature engineering (ROI-based descriptors), which is relatively uncommon and well-motivated clinically.
    • The method introduced demonstrates strong performance gains (e. g. , AUC 0.984, MCC 0.854) over both classical methods and diffusion-based baselines.
    • The authors provide a systematic ablation study, clearly showing the impact of synthetic data and the necessity of phenotype filtering.
  • 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 evaluation suffers from a limited dataset size and single-center cohort (n=92) which reduces reliability and generalizability.
    • There is no external validation dataset which is critical especially given domain shift (lighting, ethnicity, acquisition device).
    • Lack of comparison with more recent controllable diffusion methods (e. g. , MedDiff-FT [Xie et al. , MICCAI 2025], VicTr [Susladkar et al. , ICCV 2025]) that explicitly address structural and clinical consistency.
    • The phenotype descriptors are relatively simplistic (intensity ratios, Laplacian variance, gradients) and may not capture complex anatomical variations compared to learned feature-based approaches.
    • The binary classification setting collapses malnutrition severity levels, limiting clinical usefulness compared to prior works considering richer stratification (e. g. , GLIM-based grading).
    • Evaluation of synthetic image quality relies heavily on FID/AFFID, which are known to correlate poorly with clinical realism in medical imaging (as discussed in diffusion literature [Kazerouni et al. , 2023 survey]).
  • 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?

    The paper addresses an important and underexplored clinical problem and proposes a reasonably novel and well-motivated solution that combines diffusion-based augmentation with clinically informed phenotype filtering. The idea of enforcing phenotype consistency via ROI-based descriptors and Mahalanobis filtering is intuitive and empirically effective, leading to strong performance gains in a challenging few-shot setting. However, the overall contribution remains somewhat incremental relative to recent advances in controllable medical diffusion models, and the validation is limited by a very small, single-center dataset without external testing. Additionally, the handcrafted nature of the phenotype features and the binary problem formulation constrain the clinical and methodological impact. As a result, while the approach is promising and shows potential, stronger validation and positioning against more recent methods would be needed for a higher confidence acceptance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes NutriDiff, a diffusion-based data augmentation framework for facial malnutrition screening in elderly populations. The method fine-tunes Stable Diffusion using view-conditioned LoRA adapters to generate synthetic facial images and applies a phenotype-consistency filtering step based on handcrafted facial region descriptors. The generated data is then used to improve a downstream classifier (CLIP-based) for malnutrition prediction. Experiments on a small private dataset show improvements in classification performance when synthetic data is incorporated. While the approach combines diffusion models with feature-based filtering, the overall contribution primarily lies in data augmentation and computer vision modeling rather than advancing medical image computing methodologies.

  • 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 use of diffusion models with phenotype-based filtering is a reasonable attempt to improve synthetic data quality.
    • The integration with a downstream classifier shows empirical performance gains.
  • 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.Misalignment with MICCAI Scope (Major Concern). The primary limitation of this work is that it is not well aligned with the core focus of MICCAI. The paper mainly focuses on facial image-based classification and diffusion-based data augmentation, which are more typical of the computer vision community (e.g., CVPR, ICCV, ECCV). The methodological contributions are centered around generative modeling and feature filtering rather than advancing medical imaging techniques, clinical workflows, or image-based diagnosis grounded in medical imaging modalities such as MRI, CT, or ultrasound. 2.Limited Medical Imaging Contribution. Although the application is related to healthcare (malnutrition screening), the method operates entirely on 2D facial photographs, which are closer to natural image processing rather than medical imaging. The paper does not leverage domain-specific medical imaging knowledge, physics, or clinically grounded modeling, which are central to MICCAI contributions. 3.Incremental Methodological Novelty. The technical contributions are largely incremental:

    • Fine-tuning Stable Diffusion with LoRA is now standard practice.
    • The phenotype filtering step is based on handcrafted features and Mahalanobis distance, which is relatively simple. 4.Weak Clinical Validation. The experiments are solely done with a private in-house dataset. 5.Limited Innovation in Diffusion Modeling. The paper does not propose a novel diffusion formulation but rather adapts existing Stable Diffusion models with conditioning and filtering. Compared to recent advances in medical diffusion models (e.g., structure-aware, physics-informed, or uncertainty-aware diffusion), the contribution is relatively limited.
  • 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?

    The main reason for rejection is that the paper is better suited for a computer vision venue (e.g., CVPR, ICCV, ECCV) rather than MICCAI. The contributions are primarily focused on diffusion-based image generation and data augmentation for facial images, with limited advancement in medical imaging methodology or clinical relevance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    paper proposes NutriDiff, a multi-view diffusion augmentation framework for facial-based elderly malnutrition 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.

    1.The four facial ROIs (temporal, periorbital, malar, jawline) are grounded in established nutritional physiology and their use in the paper gives it strong clinical motivation. 2.The ablation study in Table 2 convincingly demonstrates that naive synthetic augmentation can degrade performance and their phenotypic filtering consistently improves all metrics 3.The multi-view design with subject-level score averaging is practical for bedside screening 4.Using separate LoRA adapters per view-class pair is a parameter-efficient strategy that avoids overfitting on the small dataset while capturing view-dependent morphological differences. 5.The paper is clearly written with a logical flow

  • 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 dataset is extremely small and from a single center. this limits generalisability 2.No Cross Validation results 3.No statistical significance tests 4.There is domain shift between train (ambulatory patients) and test(bedridden ). This has not been analysed or discussed properly 5.No sensitivity analysis of key hyperparameters such as 95th percentile Mahalanobis threshold, the generation multiplier β=8, and the synthetic loss weight γ=0.3.6.No standard augmentation baselines (geometric transforms, color jitter, CutMix, mixup) or non-diffusion generative methods (GANs, VAEs) are included 7.Individual components have limited novelty 8.Reproducibility : The dataset is private. No code release is mentioned

  • 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

    Go to the weakness section to address the comments. A t-SNE or UMAP visualization of real vs. synthetic feature distributions before and after filtering would strengthen the qualitative evidence for the filtering mechanism. What is the landmark detection failure rate on the synthetic images? This affects the practical yield of the generation pipeline.

  • 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 addresses a clinically relevant and underexplored problem. The problem statement is very important with scant literature so the lack of baselines or diverse datasets is understandable. There is lack of some analysis and tests that weaken the paper contribution along with an already weak novelty.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 the Area Chair and reviewers for their constructive comments and encouraging assessment of our manuscript. We appreciate their recognition of the clinical relevance and underexplored setting of elderly malnutrition screening (R1/R3), the phenotypic consistency filtering based on facial regions associated with nutritional status (R1/R3), the practical multi-view design (R3), and the empirical gains shown in the ablation study (R1/R3).

To clarify the relevance to MICCAI and the work’s medical contribution (R2), our goal is not generic face generation. The paper investigates phenotype-consistent diffusion augmentation for elderly malnutrition screening. Facial images may capture phenotypic cues linked to nutritional status. This potential relevance to healthcare screening and geriatric assessment motivated our MICCAI submission. The filtering step uses facial ROIs tied to nutritional status to reduce mismatch in synthetic samples and to support an image-based clinical screening task.

We agree with the reviewers that the small single-center cohort and the absence of external validation are important limitations (R1/R3). These limitations are partly due to the scarcity of data and prior work on facial image analysis for screening elderly malnutrition. To the best of our knowledge, no public dataset currently targets this task for individuals aged 65 years or older. We also clarify that the submitted test set includes bedridden patients and those receiving enteral nutrition, thereby providing an initial evaluation under more complex clinical conditions.

For cross-validation and statistical rigor (R3), Table 2 in the submitted manuscript reports subject-level 5-fold cross-validation on the training set for ablation and stability analysis. These experiments compare real data, unfiltered synthetic augmentation, and phenotype-filtered synthetic augmentation across two backbones, with the mean and standard deviation reported.

We agree that additional comparisons with traditional augmentation, GAN/VAE-based generation, and recent controllable diffusion methods would further strengthen the evaluation (R1/R3). Given the page limit and the focus of the study, the submitted experiments prioritize diffusion-based augmentation. The main objective is to assess whether phenotypic consistency filtering improves the clinical consistency and downstream screening utility of diffusion-generated samples. Broader comparisons across augmentation families will be an important direction for future work.

We also appreciate the suggestions on hyperparameter sensitivity, landmark detection failure analysis, and t-SNE/UMAP visualization of real and synthetic feature distributions (R3). These analyses would provide useful evidence for understanding the generation and filtering process. Larger multi-center validation, more rigorous statistical and hyperparameter analyses, broader baseline comparisons, and external validation will be important extensions of this work.

We again thank the Area Chair and reviewers for their thoughtful feedback. Their comments will guide our future validation, broader comparisons, and continued clinical extension of this work.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviews are generally positive, with two reviewers recommending weak acceptance and one reviewer recommending weak rejection. The paper addresses an underexplored and clinically relevant problem: facial-image-based malnutrition screening in elderly patients. I do not agree with the concern that the work is out of scope for MICCAI. Facial image diagnosis can be considered an important part of medical image analysis, with relevance to healthcare screening, geriatric assessment, and related areas such as AI-assisted facial analysis in healthcare and Traditional Chinese Medicine. The proposed diffusion-based augmentation framework, especially the phenotype-consistency filtering using clinically motivated facial regions, is well motivated and supported by ablation results.

    Given the positive overall assessment, clear clinical motivation, and empirical gains, I recommend early acceptance. However, several concerns raised by multiple reviewers should be addressed carefully in the final version, including the small single-center dataset, limited statistical rigor, absence of cross-validation or significance testing, lack of external validation, insufficient comparison with standard augmentation and recent generative baselines, and limited analysis of domain shift and hyperparameter sensitivity. The authors are strongly encouraged to clarify these issues, strengthen the discussion, and include recommended analyses where feasible so that the published work can have stronger impact.



back to top