Abstract

Mental health screening predominantly relies on subjective questionnaires, leaving observable biological signals uncaptured. The skin-brain axis, mediated by neuro-immune and endocrine pathways, links psychological states to measurable changes in facial skin phenotypes. However, no computational framework has directly modeled this link for individual-level mental health risk prediction. We propose Skin2Mind, to the best of our knowledge, the first psychodermatology-grounded framework that integrates multi-spectral facial skin images (white, cross-polarized, and UV illumination) with structured skin analysis reports covering pigmentation, pilosebaceous activity, vascular response, structural aging patterns, barrier sensitivity, and facial morphology. To address inter-modal redundancy in co-acquired skin imaging and structured report modalities, we introduce a Hierarchical Anchor Residual Fusion strategy that treats UV images as the anchor modality and integrates auxiliary modalities through gated residual pathways. Evaluated on a clinical cohort of 192 subjects, Skin2Mind achieves an Area Under the ROC Curve (AUC) of 81.66% for DSM-5-aligned at-risk mental health classification, outperforming representative tabular, unimodal, and vision-language baselines. More importantly, semantic-level Leave-One-Feature-Out analysis reveals that the model predictions are driven by dermatologic dimensions consistent with skin-brain axis mechanisms, including epidermal barrier sensitivity, periorbital fatigue cues, pilosebaceous activity, and stress-related structural aging features. These findings provide the first computational evidence that quantitative facial skin phenotypes can serve as non-invasive, scalable biomarkers for mental health risk screening.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0579_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/skin2mind/miccai-2026

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiuJia_Skin2Mind_MICCAI2026,
        author = { Liu, Jiahe AND Nhu, Duong AND Zhao, Xiangyu AND Jiang, Yiwen AND Yan, Siyuan AND Xue, Xiangyi AND Shen, Yaling AND Camargo de Oliveira, Guilherme AND Fong, Stephanie AND Xu, Qingyang AND Wang, Zimu AND Wang, Fudi AND Zhang, Xiaoqing AND Dwyer, Dominic AND Kuhlmann, Levin AND Ge, Zongyuan},
        title = { { Skin2Mind: A Multimodal Framework for Mental Health Risk Screening via Facial Images and Structured Skin Reports } },
        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

    Authors propose a method that analyses the visual appearance of patients’ facial skin to detect those patients with conditions of what authors summarize as “mental health at risk”. The proposed method leverages the idea that mental state may visually manifest in the appearance of the facial skin (“skin-brain axis”). To this end, their method fuses multi-spectral images of a patient’s face (polarized, UV, white light) with auto-generated textual “skin reports” to ultimately produce a binary classification output (“at risk” vs. “not at risk”). Input fusion is formulated as a multi-stage process, with an embedding of the UV image forming the basis and successively adding the remaining image and text representations via learned gated residual adapters. Authors evaluate their method on an in-house dataset, including an extensive set of baseline methods and ablations.

  • 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 is well-written. The proposed method is well-described and appears as reasonably novel, maybe even as pioneering, regarding work along the “skin-brain axis”. Experiments and ablations are extensive, results look convincing.

  • 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.

    Apart from the fact that authors don’t mention their intended publication of the method’s code, I only see one major weakness in the authors’ work: It is not immediately clear to me why the steps from transforming tabular skin reports to textual skin reports to text embeddings are really necessary. Since the authors use XGBoost as a tabular, text-only baseline and since the results appear quite competitive with the textual, text-only baselines, couldn’t they likewise have used XGBoost outputs as the inputs to an adapter, similar to how they currently use the text model outputs? Maybe this could have provided a more lightweight alternative, and maybe this could be discussed as an alternative approach.

  • 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 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
    • Abstract: Though admittedly pretty common, the abbreviation “AUC” should be introduced somewhere.
    • Method (Sect. 2): What is the size (in pixels) of the input images?
    • Method (Sect. 2): Did the authors consider alternative outputs other than binary (“at risk”/”not at risk”)? While I understand the use of a binary classifier for introducing the method, I doubt the output’s clinical relevance. Maybe classifications or regressions regarding more fine-grained clinical conditions could be an alternative here? This is by all means a question that I just ask out of curiosity, still maybe the authors could consider it for their Discussion or their potential future work.
    • Residual fusion (Sect. 2.2): – The text introduces f_uv as a “linear projection”, while, if I understand correctly, the corresponding blue box in Fig. 2 is labeled “MLP”. An MLP, however, is in general not a linear projection (e. g. through its activation functions). So which one is it, a linear projection or an MLP? Same for P_t. – z_text, z_fused: Since “text” and “fused” are not variables themselves, they should not be italicized.
    • Implementation details (Sect. 3.1): – What activation function is used in the MLP? – an initial learning rate is given for the optimizer, which implies to me a learning rate schedule (i. e. an adjusted learning rate over time). So either the used learning rate schedule should be added (together with its remaining choice of parameters) or the word “initial” should be dropped.
    • Baseline comparisons (Sect. 3.2): – “Encoding … with LLMs consistently outperforms … XGBoost” – I think “consistently” is an overstatement here, given that only 2 out of 6 LLMs (w. r. t. F1) or 4 out of 6 LLMs (w. r. t. AUC) actually outperform XGBoost (see Table 1). Authors should consider rephrasing. – “Statistical differences are primarily observed in AUC rather than F1, suggesting differences in ranking rather than fixed-threshold performance. “ – I do not understand this sentence. In particular, how does AUC reflect “differences in ranking” (what is the latter supposed to mean in the first place), while F1 should reflect “fixed-threshold performance” (again, what is the latter supposed to mean in the first place)? Authors should perhaps consider rephrasing. – “The advantage of multimodal integration lies in … computational efficiency. “ – How does multimodal integration provide efficiency; in particular, as compared to what? Authors should consider a clarification. – “Compared with end-to-end VLMs such as Qwen2.5-VL (AUC std ± 10.20%), HARF demonstrates more stable behavior (AUC std ± 4.32%)” – Did the authors intentionally not mention Qwen3-VL here, which to my understanding of the authors’ definition of stability, actually shows a more stable behavior than HARF (AUC std ± 3.86%)? If so, what was the reason for this nonmention? Authors should consider adapting their corresponding observation.
  • 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    As already mentioned above, I think the method is reasonably novel (even pioneering perhaps), the evaluation is sound and extensive, and the results are convincing. I am pretty positive that authors can refute my only major concern via an extended discussion or, if necessary, via rebuttal.

  • 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



Review #2

  • Please describe the contribution of the paper

    The authors propose Skin2Mind, which they position as the first multimodal framework for mental health risk screening based on facial skin images and structured skin reports. The method combines modern domain-specific dermatology foundation models for image representation learning with LLM-based encoding of structured skin reports. To integrate the multimodal embeddings, the authors introduce a Hierarchical Anchor Residual Fusion (HARF) strategy, which aims to reduce inter-modal redundancy and stabilize fusion by treating UV facial images as the anchor modality and incorporating the remaining modalities through gated residual pathways.

  • 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 explores a novel and unconventional application setting. The use of facial skin phenotypes and structured skin reports for mental health risk screening is, to my knowledge, highly original and makes the paper stand out from more conventional multimodal medical AI studies. The work is interesting because it attempts to test whether current foundation models can capture the relationship between objective facial skin phenotypes and individual-level mental health risk.

    2) the paper proposes a new multimodal fusion mechanism, HARF, which is a meaningful technical contribution. The hierarchical anchor-residual design is well motivated by the fact that the modalities are partially redundant, since the images and reports are derived from the same acquisition pipeline. 3) the paper makes good use of modern pretrained backbones and includes a reasonably broad comparison of unimodal and multimodal baselines, including vision-only, text-only, and VLM-based approaches. The inclusion of ablation studies across fusion strategies and modality choices strengthens the empirical section.

  • 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.

    A major weakness is that the biological and clinical justification of the core hypothesis remains somewhat limited. The paper makes an ambitious claim about the relationship between skin phenotypes and mental health risk, but the supporting biological motivation is not yet fully convincing. In my view, this type of claim would benefit from stronger support from more established biomedical and clinical literature.

    A second weakness is the small and imbalanced dataset. The study is conducted on only 192 subjects, with a substantial class imbalance, which raises concerns about the robustness and generalizability of the findings. Given the novelty and sensitivity of the proposed application, stronger validation on a larger and more diverse cohort would be important.

    Another concern is that the paper criticizes conventional mental health screening for relying on self-reported questionnaires and clinical interviews as subjective tools, but the proposed labels are themselves based on DSM-5-derived questionnaires. This weakens the claim that the method provides a substantially more objective alternative, since the supervisory signal still originates from questionnaire-based assessment.

  • 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 rate this paper 4/6 because it is novel, ambitious, and methodologically interesting, and it presents a creative use of modern foundation models to test a new hypothesis in medicine. In particular, the paper is strong in its use of multimodal learning, its proposed fusion mechanism, and its comparison with several relevant baselines. The ablation studies are also helpful and suggest that the proposed fusion design contributes meaningfully beyond simpler alternatives.

    At the same time, I do not score it higher because the paper makes a relatively strong clinical and biological claim on the basis of a small, imbalanced cohort and labels that are still questionnaire-derived. Thus, while the technical contribution is interesting and the experimental study is promising, the current evidence is not yet sufficient to fully support the broader implications suggested by the paper. I would have liked to see stronger biological grounding, validation on a larger dataset, and a more clinically detailed evaluation setting.

  • 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

    The authors present Skin2Mind, a Hierarchical anchor residual fusion (HARF) strategy to integrate multi-spectral facial skin images and skin analysis reports for the prediction of Mental Health at risk. Authors claim that Skin2Mind is the first end-to-end computational framework in this classification task.

  • 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 authors perform an appropriate evaluation of independent modalities with respect to state-of-the art computational methods for imaging, text and VL. Authors analyzed different text models in the prediction of skin reports. And also considering two imaging models: a ResNet50 network and the imaging backbone of DermLIP, a VL foundation model. In the VLM configuration, they compared with Qwen3-8B-VL model.

    • Also comparison using Naive Imaging-text fusion demonstrated the advantage of the HARF method.

    • The interpretability from the feature importance results, provides an additional value of the work, given that commonly VLMs lack explainability to which words or textual patterns lead models to the good performance.

    • The different part of the method are well explained and the pipeline of the proposed integration is easy to follow. Also results section is well organized and support the selection of the proposed configuration.

  • 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.
    • Given that the authors used the DermLIP imaging encoder, why to not use also the DermLIP text encoder to adapt their VL representations ? Is not clear the chosen of Qwen2.5-7B for the text processing.

    • Why the authors Qwen2.5-7B for the text processing if Mistral-7B had a better performance for text-only? How authors argue that HARF improve Qwen2.5-7B in the HARF-Qwen method, but HARF doesn’t improve Mistral in the HARF-Mistral method? Why HARF improves the performance when using specifically Qwen and not different text encoders ?

    • Why authors compared only with Qwen-VL for the VLMs ? and for example didn’t consider to fine-tune DermLIP VL method ?

    • Authors should provide better arguments on the residual feature approach which is the main block of HARF. Why it is designed like that ? Why the input is the Xm Xuv and the difference ?

  • 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 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
    • The manuscript has minor errors. For instance, in the Fig. 1.in the White adapter, You have 3 inputs, but Xwhite is repeated. I think the one in the bottom should be X_uv.

    • The last paragraph of section 3.2 The AUC std values are missing, e.g “… (AUC std +- 10.2%).

  • 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?
    • Although should be interesting a better comparison with state of the art VLMs or at least with the foundation VL model on dermatology, their method shows a superior performance and coherence from the others results (ablations of the method, and analysis of features relevance).
  • Reviewer confidence

    Very confident (4)

  • [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 thank the reviewers and Area Chairs for their constructive feedback and for recognizing the novelty and interdisciplinary potential of Skin2Mind. We have carefully revised the camera-ready version to improve clarity, reproducibility, and alignment between the manuscript text and the implemented framework, while keeping all experiments, numerical results, and conclusions unchanged.

First, following reviewer suggestions regarding framing and interpretation [R2], we refined several statements throughout the manuscript to avoid overstating “objectivity” in mental health assessment. We clarified that DSM-5 labels remain questionnaire-derived clinical measures and revised the Introduction and Abstract to position Skin2Mind as a complementary computational screening framework based on quantitative facial skin phenotypes rather than a replacement for standard psychiatric assessment. We also refined the discussion comparing AUC and F1 to clarify the distinction between ranking performance and fixed-threshold classification behaviour [R1].

Second, we improved methodological clarity in Section 2.Reviewer comments correctly noted that several textual descriptions were inconsistent with the architecture illustrated in Fig. 1 [R1]. We therefore updated the manuscript so that the written description precisely matches the implementation used during training and evaluation. Specifically, projection and gate components previously described as “linear layers” are now correctly described as two-layer MLP modules, consistent with the implemented method shown in Fig. 1.Additionally, the description of the text-side fusion component has been clarified for consistency with the implementation. We also added a brief rationale for using Qwen2.5-7B instead of the DermLIP text encoder and conventional tabular models such as XGBoost for long structured reports [R1, R3]. Our skin reports contain more than 400 heterogeneous descriptors with compositional semantic relationships across pigmentation, structural aging, barrier sensitivity, and pilosebaceous activity. These reports exceed the short-caption scope of CLIP-style text encoders, while tabular models treat features largely independently and cannot fully capture higher-order semantic interactions across descriptors. We therefore used frozen LLM embeddings to obtain richer semantic representations of the structured reports. These are documentation clarifications only: no code, hyperparameters, datasets, or experiments were modified, and all reported results (including the headline AUC of 81.66%) remain identical to the accepted version.

Third, we addressed reproducibility and presentation concerns raised by the reviewers [R1, R3]. We added the public code repository, clarified implementation details (e.g., constant learning rate, ReLU activations), corrected several notation inconsistencies and figure labels, and revised parts of the baseline discussion to make comparisons more precise and balanced. We also added the required ethics approval, informed consent, acknowledgments, disclosure statements, and final de-anonymized author information in accordance with the Springer LNCS camera-ready guidelines.

Overall, all revisions are minor clarifications, wording refinements, or mandatory camera-ready updates. No new experiments were added, no results were changed, and the scientific conclusions of the paper remain unchanged.

Thank you again for the valuable feedback and for supporting this interdisciplinary direction at the intersection of multimodal AI, dermatology, and mental health screening.




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.

    All reviewers have reached a consensus to accept the paper.



back to top