Vitiligo state assessment based on progressive transfer learning and multimodal domain adaptation

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Vitiligo is a common skin depigmentation disorder requiring accurate assessment for effective treatment.
• Current data collection for vitiligo assessment is complex, costly, and results in limited dataset sizes, hindering deep learning model performance.
• Transfer learning in medical imaging faces challenges due to feature discrepancies between natural and medical images and poor generalization across different data modalities.

Purpose of the Study:

• To develop an accurate vitiligo state assessment method using progressive transfer learning and multimodal domain adaptation.
• To overcome limitations of conventional transfer learning in medical image analysis, particularly for vitiligo.
• To enhance the generalization and accuracy of deep learning models for dermatological conditions with limited labeled data.

Main Methods:

• Employed progressive transfer learning with multi-step fine-tuning on unlabeled medical images to bridge the gap between natural and medical image features.
• Utilized an adaptive parameter unfreezing strategy for labeled target data to improve model adaptability and accuracy.
• Integrated a multimodal domain adaptation approach using CycleGAN and HSV color space transformation to mitigate modality-specific feature impacts.

Main Results:

• The proposed method significantly improved accuracy compared to conventional transfer learning.
• Achieved a 2.5% accuracy improvement on the clinical modality dataset.
• Achieved a 3.2% accuracy improvement on the Wood's lamp modality dataset.

Conclusions:

• The developed progressive transfer learning and multimodal domain adaptation method offers accurate vitiligo state assessment.
• This approach effectively addresses challenges associated with limited labeled data and modality differences in medical image analysis.
• The method has potential applications for various multimodal dermatological diseases requiring accurate assessment with scarce labeled data.