An Intelligent Diagnostic Model for Melasma Based on Deep Learning and Multimode Image Input

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melasma diagnosis relies on subjective physician assessment, posing challenges for inexperienced practitioners and potentially leading to incorrect treatments.
• Accurate melasma diagnosis is crucial for effective treatment and patient outcomes.

Purpose of the Study:

• To develop and validate an intelligent diagnostic system utilizing deep learning for improved melasma image analysis.
• To enhance the accuracy and reliability of melasma diagnosis compared to traditional methods.

Main Methods:

• A dataset of 8010 VISIA system images (4005 melasma, 4005 non-melasma) was used for training and testing deep learning models.
• Evaluated DenseNet, ResNet, Swin Transformer, and MobileNet architectures for binary classification of melasma.
• Investigated the impact of fusing multiple image modes (e.g., NORMAL, BROWN SPOTS, UV SPOTS) on diagnostic performance.

Main Results:

• The DenseNet121-based network achieved 93.68% accuracy and 97.86% AUC for melasma classification.
• Gradient-weighted Class Activation Mapping confirmed the interpretability of the diagnostic model.
• Combining 'NORMAL,' 'BROWN SPOTS,' and 'UV SPOTS' modes yielded the highest performance: 97.4% accuracy and 99.28% AUC.

Conclusions:

• Deep learning models are effective for diagnosing melasma using clinical images.
• The proposed network demonstrates excellent performance and high accuracy, particularly when utilizing multiple VISIA image modes.