List of Papers Browse by Subject Areas Author List
Abstract
Psoriasis and seborrheic dermatitis (SD) often present with highly similar clinical features such as erythema and scaling, leading to frequent diagnostic ambiguity, particularly in primary care and teledermatology. Existing AI approaches based on mobile RGB imaging have shown limited discrimination performance because the two conditions can exhibit overlapping visual manifestations. To address this challenge, we propose a smartphone-based portable multimodal optical imaging and analysis framework that acquires co- and cross-polarized multispectral reflectance images and ultraviolet (UV)-excited autofluorescence, together with a modality-aware deep learning model, mSpecFusion-Net, designed for robust cross-modality integration. The proposed model encodes each modality in parallel and employs an iterative fusion Transformer that repeatedly aggregates complementary information across modalities to form a unified representation for differential diagnosis. Across patient-wise splits, mSpecFusion-Net (PCA + iterative fusion) delivers the strongest overall performance (F1 = 0.9057, ROC-AUC = 0.9461), surpassing RGB-only baselines and widely used backbones. This demonstrates that explicitly reducing inter-wavelength redundancy and progressively refining cross-modality evidence is critical for separating clinically similar conditions. The proposed smartphone multimodal pipeline is readily deployable for teledermatology and scalable to other multimodal mobile diagnostics.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4806_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Seong-Hyun-0224/mSpecFusion-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KimSew_mSpecFusionNet_MICCAI2026,
author = { Kim, Sewoong AND Kim, Seong-Hyun AND Cavalcanti, Thiago Coutinho AND Lee, Jonghun AND Lee, Moon Hwan AND Lee, Donghun AND Hwang, Jae Youn},
title = { { mSpecFusion-Net: Intelligent Multimodal Smartphone Imaging for Mobile Differential Diagnosis of Psoriasis and Dermatitis and Deep Learning } },
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 presents a hardware-software co-design framework for differentiating between two visually similar skin conditions: Psoriasis and Seborrheic Dermatitis (SD). The authors propose a custom smartphone-based multimodal dermascope (capturing co-polarized multispectral, cross-polarized multispectral, and UV-excited autofluorescence images) paired with a novel deep learning architecture (mSpecFusion-Net). The model utilizes PCA for spectral compression and an iterative-fusion Transformer to aggregate cross-modal features.
The integration of custom mobile optical hardware with a tailored multimodal deep learning pipeline is quite compelling and addresses a real clinical need in teledermatology, even though the sample size is slightly on the lower side (32)? The systematic ablation studies (Tables 2 and 4) are well-structured and clearly justify the architectural choices. However, crucial methodological details regarding data processing and evaluation metrics are missing (especially in Tables 3 and 5), which must be addressed before publication. Also, crucial comparisons with recent work on missing, which should demonstrate how mSpecFusion-Net improves over SOTA architectures and how iterative fusion improves specifically over modern fusion methods.
- 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.Innovative Hardware-Software Integration: The combination of an 8-band LED multiplexer with polarization filters and UV autofluorescence on a mobile device seems highly innovative. It provides a unique solution to the limitations of standard RGB imaging.
2.Rigorous Validation Strategy: The use of patient-wise 5-fold cross-validation is excellent. In dermatology AI, random splitting often leads to severe data leakage (where the model learns the patient’s background skin rather than the disease). The authors correctly avoided this pitfall.
3.Strong Ablation Studies: The paper does an excellent job of justifying its complexity. Table 2 proves the utility of PCA, Table 3 validates the iterative fusion module over standard early/late fusion (even though there are concerns regarding this table), and Table 4 confirms the necessity of the multimodal hardware setup.
4.Clinical Relevance: Targeting Psoriasis vs. Seborrheic Dermatitis is a highly practical objective, as these are notoriously difficult to distinguish in primary care and RGB teledermatology.
5.Architectural Innovation: The use of PCA, the concept of iterative fusion, and the overall design seem sufficiently novel and well motivated.
- 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 major weakness of this paper is the lack of clarity and incompleteness when it comes to the evaluations against SOTA architectures and fusion methods. I am listing them below:
1.Ambiguity in Evaluation Metrics (Patch vs. Patient Level): Section 2.4 states that images are “partitioned into a fixed grid of nonoverlapping patches” and that these patches are used as training instances. However, the evaluation tables (Tables 2-5) do not clarify if the reported F1, Accuracy, and AUC metrics are evaluated at the patch-level, image-level, or patient-level.
Why this matters: If a model predicts 100 patches from a single patient’s image and gets 90 right, patch-level metrics will look fantastic, but clinically, we only care about the single image/patient diagnosis. The authors must report patient-level or image-level aggregated metrics (e. g. , via majority voting across patches).
2.Missing PCA Parameters: The paper mentions applying PCA to the multispectral stacks to reduce inter-wavelength redundancy, but omits critical reproducibility details. How many principal components were retained? What percentage of variance was explained by the retained components?
3.Unfair Baseline Comparisons: In Table 5, the proposed model is compared to standard backbones. It is unclear how these baselines processed the multimodal data. Did they receive a concatenated tensor of all modalities (early fusion)? If these baselines were only given RGB images, or if they were not adapted to handle the high-dimensional multispectral stacks properly, the comparison to mSpecFusion-Net is an apples-to-oranges comparison. It doesn’t add any value at all. How these baselines were attained needs critical attention.
4.Fusion Methods: Similarly, Table 3 mentions fusion methods, but fails to provide any detail whatsoever on the specifics (late fusion, how? Concatenation? Attention? Gated?)
5.Lack of comparison to recent works: The authors fail to compare their work against recent multimodal, multispectral, and fusion-based studies. For example, the following papers could be discussed to outline the improvements: i) https://doi. org/10.1016/j. ibmed. 2025.100271 ii) https://doi. org/10.1109/JBHI. 2026.3682634
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I am leaning towards accepting because the authors addressed a relevant clinical problem, and the proposed framework and architecture are quite innovative. Weak confidence because of the lack of clarity in the evaluation, as mentioned. However, if the authors can discuss/compare some fusion-based models and clarify the details for Tables 3 and 5 properly, I believe this paper should be accepted.
- 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
Review #2
- Please describe the contribution of the paper
This paper presents a smartphone-based multimodal imaging and artificial intelligence framework for the differential diagnosis of psoriasis and seborrheic dermatitis in mobile and remote dermatology settings. The authors combine a portable multimodal dermoscope with a modality-aware deep learning model that performs parallel modality encoding and iterative cross-modality fusion. In addition, the study introduces a PCA-based spectral compression strategy to reduce redundancy across wavelengths before fusion. The results show that the proposed framework outperforms RGB-only, single-modality, conventional fusion, and several generic deep learning architectures, suggesting that clinically grounded multimodal acquisition and progressive cross-modality reasoning can improve the discrimination of visually similar inflammatory skin diseases.
- 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.
Psoriasis and seborrheic dermatitis often present with similar visual appearances, making research on mobile screening and remote dermatological diagnosis highly clinically relevant. This paper proposes a coherent end-to-end framework that includes a smartphone-based multimodal dermoscope, acquisition and calibration procedures, and a dedicated multimodal analysis model. As a result, its contribution is more practically meaningful than that of a purely retrospective classification study. The proposed iterative fusion design is a meaningful innovation. It allows the model to progressively exchange and refine cross-modal evidence. The results suggest that this strategy outperforms conventional fusion approaches. The full multimodal iterative fusion model performs clearly better than the RGB-only input setting, demonstrating that explicitly modeling multimodal complementarity is more effective than simply extending generic architectures.
- 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 method used in this paper integrates parallel modality encoders, Transformer-based fusion, PCA-based spectral compression, and iterative multimodal fusion. While the use of these components is reasonable, the work does not demonstrate a strongly novel algorithmic advance. In my view, the most distinctive contribution of this paper lies in the multimodal mobile diagnostic workflow rather than in a fundamentally new core algorithmic design. During training, the authors divide the images into non-overlapping patches and assign patient-level diagnostic labels to each patch. I am concerned that this strategy may reduce lesion-level contextual information and may not fully reflect how dermatologists integrate the overall morphology of a lesion in clinical practice. I would encourage the authors to provide more discussion to justify this design choice. Given that the goal of this work is mobile-device-based diagnosis, the paper should report more information on acquisition time, inference time, computational cost, and operational complexity. I encourage the authors to more fully quantify the practical burden of the entire multimodal workflow in order to demonstrate the deployability of the proposed framework. The presentation of symbols and formulas in the manuscript is not sufficiently clear. For example, the meaning of F in Figure 1 is not explained in the text. I encourage the authors to improve these aspects to enhance the overall quality of the manuscript.
- 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?
This paper addresses a clinically relevant and practically important diagnostic problem, and proposes a coherent system-level solution. The combination of smartphone-based multimodal acquisition, PCA-based spectral compression, and iterative modality-aware fusion is well motivated, and the experimental design is generally sound. The results show clear improvements over RGB-only input, single-modality settings, standard fusion strategies, and several generic deep learning architectures, which supports the value of the proposed application framework. That said, the work is still limited by relatively modest methodological novelty at the core algorithmic level and by the lack of details regarding practical deployment. Overall, I believe this study has strong application potential and, in its current form, is above the acceptance threshold.
- 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 paper presents a smartphone-based multimodal imaging framework (DermaSpecFusion) and a modality-aware iterative-fusion deep learning model, mSpecFusion-Net, specifically designed for differential diagnosis of psoriasis and seborrheic dermatitis. Their framework includes acquisition of complementary optical contrasts (co-/cross-polarized multispectral reflectance and UV-excited autofluorescence); PCA-based spectral compression to reduce inter-wavelength redundancy before fusion, and iterative cross-modality fusion. Experiments are done using 2032 multimodal images collected from 32 patients (14 SD, 18 psoriasis).
- 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.
– Paper is well written. Motivation is well presented. – The proposed mSpecFusion-Net architecture, using iterative fusion, seems to be a novel approach for integrating complementary information across modalities. The results presented indicate that this improves the robustness compared to simpler fusion methods – Considering 5-fold cross validation in experiments for comparative analysis – Considering and addressing the issue of inter-wavelength redundancy using PCA for improved performance
- 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.
- Limited novelty as there has been Smartphone-based multispectral imaging systems (mentioned by authors as well).
- Comparison with traditional 2019-2021 DL methods only (ResNet, ViT, efficientNet, etc)
- No discussion of statistical significance results and standard deviation of 5-fold cross validation.
- No computational complexity related discussion
- Reproducibility concerns. No discussion by authors regarding providing codes and dataset, or making these available later.
- Differences with paper [6] not clearly highlighted. Following related paper not considered (Wang et al., “Enhancing Diagnosis of Psoriasis and Inflammatory Skin Diseases: A Spatially Aligned Multimodal Model Integrating Clinical and Dermoscopic Images”, Journal of Investigative Dermatology, Volume 145, Issue 11, November 2025)
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
It is not clear if ethical approval including informed consent of patients was taken or not. There is no such description of these in the paper. Authors only mention the following: “We collected multimodal image sets from 32 patients (14 SD, 18 psoriasis), totaling 2,032 multimodal images. Diagnoses were determined by board-certified dermatologists. We use patient-wise 5-fold cross-validation to prevent identity leakage and further to obtain reliable generalization estimates [14, 15]. The F1, accuracy, AUROC, and AUPRC were utilized as a evaluation metrics. [16, 17].”
- 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
Typos in papers. For e.g. Page 2, 2nd last para: “deelp learing model”. Not sure if it was intentional. Grammar and Typos be kindly rechecked.
- 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?
Comparison with traditional methods and not exactly with recent state of the art methods, Statistical significance analysis missing, Limited novelty and generalizability, considered limited dataset only.
- 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 all reviewers and the meta-reviewer. We address the main concerns below; minor points will be addressed in the camera-ready revision. Evaluation and cross-validation (R1, R2, Meta). All metrics in Tables 2–5 are at the patch level under strict patient-wise 5-fold CV: 32 patients are split into 10 disjoint patient-level groups; per round, 2 groups (~20%) form the test set, 1 validation, 7 training. The test partitions are disjoint and cover all patients, so each appears in exactly one test set, with no identity leakage. Patch predictions can be aggregated by averaging or voting at the lesion/patient level for deployment. PCA (R1, Meta). PCA is used as discrete band selection on calibrated multispectral stacks, not as a learned feature projection. We rank the 8 bands by cumulative absolute PCA loading and retain the top K=4 per polarization stream. Since this is discrete index selection rather than data-dependent transformation, it is performed once on the calibrated dataset and fixed across folds; parameter learning remains strictly fold-independent, so the leakage pattern typically associated with global preprocessing does not apply. Statistical reporting (R3). Across 5 folds, the model achieves F1 = 0.9057 ± 0.0612, Acc = 0.9068 ± 0.0576, ROC-AUC = 0.9461 ± 0.0540, PRC-AUC = 0.9358 ± 0.0666 (mean ± SD, micro). One fold yields F1 = 0.8174, reflecting the modest cohort; larger multi-site validation is future work. Computational cost (R3, Meta). Acquisition takes ~20 s/patient. The model has 4.18 M parameters (16.0 MB, fp32) and 44.7 GFLOPs/image (bs=1), with GPU latency 17.3 ms/image (median, 100 runs). Inference is negligible vs. acquisition, so point-of-care latency is dominated by acquisition. Ethics (R3). The study was approved by the IRB of Seoul National University Hospital (No. 1908-161-1059), with informed consent from all participants. We will add this to the clinical dataset section. Baseline fairness (R1). All baselines in Table 5 receive the same multimodal input (Co-P + Cross-P MSI with PCA + UV-AF) under the same split and protocol; RGB-only is in Table 4.We will clarify this in the Table 5 caption. Comparison with related work (R1, R2). Yang et al. and Siddiqui et al. (S3F-Net) maximize information from single-modality images via anatomical metadata and Fourier features. Our work fuses modalities from distinct physical principles, requires no metadata, and enables acquisition by non-specialists. Wang et al. uses dermoscopic and clinical images, both RGB; our multi-physical-modality approach is fundamentally different. We will add this to the Introduction. Use of recent baselines (R1, R2). We additionally tested ConvNeXt V2-Base and ConvNeXt V2-Large using the same training protocol and input setting. Both models underperformed our proposed method: ConvNeXt V2-Base achieved an F1 score of 0.736 ± 0.187, and ConvNeXt V2-Large achieved 0.567 ± 0.171, whereas our method achieved 0.906 ± 0.061.Moreover, their per-fold standard deviations (~0.17–0.19) were approximately three times higher than that of our method (0.061). Novelty (R1, R2). The key contribution is the development of an intelligent mobile multimodal framework for robust smartphone-based skin lesion assessment. Rather than relying on simple multimodal concatenation, conventional early/late fusion, or generic backbone scaling, our framework combines PCA-based spectral compression with iterative cross-modal fusion to progressively refine complementary optical evidence from co-polarized, cross-polarized, multispectral reflectance, and UV autofluorescence modalities. This model-level integration enables compact spectral representation, effective use of modality-specific and cross-modal cues, and more stable learning in a modest-sized clinical cohort. Code will be released upon publication. Typos and notation (R3). We will correct typos (e.g., “deelp learing”), clarify notation, and refine the technical descriptions throughout.
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.
Given the overall positive reviews, I recommend acceptance. Please address the following concerns in the final version: (1) clarify evaluation protocols, particularly whether metrics are reported at patch-, image-, or patient-level; (2) include missing methodological details; and (3) discuss practical deployment aspects such as computational cost and workflow feasibility. Addressing these points will be important to confirm the strength of the contribution.