Multi-Spectral Skin Disease Detection System Based on Multi-Model Multi-Layer Networks

Yi Zou, Guangming Yang, Qi Zhang, Shijie Deng, Xuanke TongPhotonics 2026 · DOI

TL;DR

This paper introduces a novel multi-spectral skin disease detection system that leverages a multi-model, multi-layer network architecture to accurately differentiate visually similar skin conditions like acne and eczema. By fusing information from multiple spectral bands and employing a hierarchical classification approach, the system achieves high diagnostic accuracy and improved interpretability, addressing key limitations in current dermatological diagnostics.

What Problem Does This Paper Address?

Skin diseases are incredibly common, affecting billions globally, and early detection is crucial for effective treatment. However, current clinical methods often suffer from subjectivity, complexity, and a lack of interpretability in diagnostic algorithms. Furthermore, much of the existing research and benchmark datasets are biased towards Caucasian skin, neglecting the spectral variations in other populations. Specifically, distinguishing between visually similar skin conditions like acne and eczema, which have high misdiagnosis rates, remains a significant challenge. This paper aims to bridge this gap by developing a more accurate, interpretable, and broadly applicable system for skin disease detection.

Methods

The core of this work is a multi-model, multi-layer network fusion framework. This system comprises two main components: a multi-spectral image acquisition device and a sophisticated diagnostic algorithm.

The acquisition device uses a circular array of LEDs emitting light at nine specific wavelengths (370 nm to 810 nm), chosen based on the absorption spectra of key skin chromophores (melanin, oxyhemoglobin, deoxyhemoglobin) and their relevance to common skin diseases. This device captures multi-spectral images of skin lesions.

The diagnostic algorithm employs a three-level hierarchical architecture:

  1. Sub-Classifier: For each of the nine spectral bands, pre-trained deep CNN architectures (ResNet50, EfficientNetB0, and RegNetX002) are fine-tuned to act as wavelength-specific sub-models. This allows for the extraction of unique features from each spectral band.
  2. Stacked Classifier: The prediction probability matrices from all sub-classifiers (9 wavelengths x 3 architectures) are concatenated. This combined feature set is then fed into a stacked classifier, which uses a combination of classifiers (Random Forest, SVM, Logistic Regression, Neural Network) to learn from these fused features.
  3. Meta-Classifier: The output of the stacked classifier is further processed by a meta-classifier (Random Forest) to produce the final diagnosis.

This multi-layer fusion approach aims to enhance interpretability by analyzing contributions at each stage and improve accuracy by combining diverse spectral and model-based features.

Datasets

The study utilized a custom-collected dataset of 2114 validated multi-spectral skin disease samples obtained from a hospital clinical setting. This dataset included:

  • 386 cases of acne
  • 235 cases of eczema
  • A control group of 200 multi-spectral images of normal skin.

The dataset covered a wide age range (1 to 93 years) and was specifically collected from an East Asian cohort, addressing the limitation of Caucasian-centric datasets. For training, the dataset was balanced by under-sampling the acne class to match the eczema class size (235 cases each), and then split into training (70%) and testing (30%) sets.

Key Insights and Ideas

  • Spectral Signatures of Similar Diseases: Even visually similar skin conditions like acne and eczema exhibit distinct spectral absorption patterns, particularly in the blue and green wavelengths, due to differences in chromophore content (porphyrins in acne, hemoglobin in eczema). This highlights the power of multi-spectral imaging for fine-grained differentiation.
  • Hierarchical Fusion for Accuracy and Interpretability: The multi-model, multi-layer network fusion approach is not just about boosting accuracy. By analyzing the outputs at each stage (sub-classifier, stacked classifier, meta-classifier), the system offers a degree of transparency into its decision-making process, which is crucial for clinical adoption.
  • Cross-Architecture Fusion Outperforms Single Architecture: Combining features from different CNN architectures (ResNet50, EfficientNetB0, RegNetX002) through fusion significantly improves performance compared to using any single architecture alone. This suggests that different architectures capture complementary information.
  • Importance of Wavelength Selection: The careful selection of nine specific wavelengths, informed by chromophore absorption, is critical for capturing the subtle spectral differences that distinguish various skin conditions.

Key Results

The proposed multi-model multi-layer network fusion algorithm achieved a diagnostic accuracy of 0.8622 and an AUC-ROC value of 0.9143 on the test set.

  • Sub-classifier performance: ResNet50 showed the best average accuracy (0.7393) and AUC-ROC (0.8255) among the three CNN architectures across all wavelengths.
  • Fusion benefits: Fusion within a single architecture (e.g., ResNet50) improved accuracy by up to 16.4% and AUC-ROC by up to 13.48%.
  • Cross-architecture fusion: Further fusion across architectures yielded even greater improvements, with accuracy increasing by up to 5.98% and AUC-ROC by up to 3.85% compared to single-architecture fusion.
  • Final meta-classifier: The final meta-classifier, after integrating outputs from multiple classifiers, achieved the highest performance metrics.

Limitations and Future Work

  • Dataset Size and Diversity: While the custom dataset is valuable, the authors acknowledge the need to expand the scale and diversity of the dataset. Future work will involve collaborating with more clinical institutions across different regions to collect data from diverse races, skin colors, and geographical areas.
  • Generalizability: The current study focuses on common skin diseases with high visual similarity. Future research aims to expand the system's coverage to a wider range of disease categories and potentially explore multimodal fusion technologies (e.g., combining spectral data with other imaging modalities) to further enhance clinical application value.
  • Interpretability: While the hierarchical fusion approach enhances interpretability, further work could explore more advanced explainability techniques to provide even clearer insights into the model's diagnostic reasoning.

Summary updated

Abstract

Skin diseases are one of the most common diseases, affecting approximately 1.9 billion people worldwide, and early detection of skin diseases can help avoid deterioration and improve treatment outcomes. Existing clinical detection methods have drawbacks such as high subjectivity and complexity, as existing diagnostic algorithms pay less attention to interpretability, and research mainly focuses on Caucasian skin. Therefore, we designed a multi-spectral skin disease detection system based on multi-model multi-layer networks to accurately differentiate similar skin diseases. The results show that the system has great potential in assisting the detection of skin diseases and can effectively improve the diagnostic accuracy and medical efficiency of skin diseases.
BibTeX
@inproceedings{zou2026multi,
  author = {Yi Zou and Guangming Yang and Qi Zhang and Shijie Deng and Xuanke Tong},
  title = {{Multi-Spectral Skin Disease Detection System Based on Multi-Model Multi-Layer Networks}},
  booktitle = {Photonics},
  year = {2026},
  doi = {10.3390/photonics13020177},
}
Figure