Multimodal Masked Autoencoder Based on Adaptive Masking for Vitiligo Stage Classification.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Vitiligo staging is complex, posing challenges for accurate diagnosis.
• Multimodal skin images (clinical and Wood's lamp) offer complementary data for vitiligo classification.
• Deep learning models struggle with multimodal data annotation and scarcity.

Purpose of the Study:

• To develop a novel deep learning model for accurate vitiligo stage classification.
• To address limitations in multimodal data annotation and scarcity for vitiligo image analysis.
• To enhance the extraction of features from multimodal dermatological images.

Main Methods:

• Proposed a Multimodal Masked Autoencoder (Multi-MAE) with adaptive masking.
• Employed an image reconstruction task to reduce reliance on annotated data.
• Utilized a pre-training strategy to mitigate multimodal data scarcity.

Main Results:

• Achieved 95.48% accuracy in vitiligo stage classification on unlabeled data.
• Demonstrated significant performance improvements over existing models like MobileNet, DenseNet, VGG, ResNet-50, BEIT, MaskFeat, SimMIM, and MAE.
• Validated the model's effectiveness in distinguishing between stable and active vitiligo stages.

Conclusions:

• The Multi-MAE model effectively classifies vitiligo stages using multimodal images.
• This approach successfully overcomes challenges related to data annotation and scarcity.
• The model shows promise as a clinical tool for evaluating vitiligo severity.