Consistent representation via contrastive learning for skin lesion diagnosis.

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Image Analysis
• Machine Learning for Healthcare

Background:

• Skin lesions, including melanoma, are common and require early detection.
• Artificial intelligence (AI) shows promise for early skin lesion detection.
• Integrating multi-modal AI in clinical settings faces challenges, including disregarded modal bias.

Purpose of the Study:

• To introduce a multi-modal feature learning technique for dermatological diagnosis.
• To address and mitigate modal bias in AI-driven skin lesion analysis.
• To develop a method for consistent representation learning across different data modalities.

Main Methods:

• A novel technique, Contrast-based Consistent Representation Disentanglement, was developed.
• Adversarial domain adaptation was used to disentangle features and create a shared representation.
• Contrastive learning was employed to ensure uniformity of common lesion attributes across modalities.

Main Results:

• The proposed technique achieved 76.1% average accuracy in multi-classification tasks.
• It surpassed existing state-of-the-art methods in performance.
• The approach successfully tackled modal bias, creating a consistent representation of lesion appearances across modalities.

Conclusions:

• A multi-modal feature learning strategy was proposed for enhanced dermatological diagnosis.
• The method demonstrated superior performance compared to other state-of-the-art approaches.
• This technique has the potential to significantly improve diagnostic precision for skin lesions.