Class-aware multi-source domain adaptation algorithm for medical image analysis using reweighted matrix matching...

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Computer Vision

Background:

• Multi-source domain adaptation (MSDA) enhances transfer learning using multiple data sources, beneficial for complex medical applications.
• Existing MSDA methods often assume identical class distributions, neglecting the critical issue of class shift in real-world medical data.

Purpose of the Study:

• To propose a novel Class-Aware Multi-Source Domain Adaptation algorithm based on a Reweighted Matrix Matching strategy (CAMSDA-RMM).
• To address the challenge of class shift in MSDA for improved medical image analysis.
• To enhance positive transfer effects and optimize source domain contributions.

Main Methods:

• Developed a class-aware strategy to strengthen positive transfer between similar classes.
• Applied first-order and second-order moment matching for effective source and target domain alignment.
• Implemented an adaptive weighting mechanism to optimize the contribution of each source domain.

Main Results:

• The proposed CAMSDA-RMM algorithm demonstrated superior performance in classification accuracy and domain adaptability.
• Experimental validation on four public chest X-ray datasets confirmed the effectiveness of the method.
• The approach successfully mitigates the impact of class shift in multi-source domain adaptation settings.

Conclusions:

• CAMSDA-RMM offers a robust solution for multi-source domain adaptation in medical imaging, particularly when class distributions differ.
• The proposed method improves the reliability and accuracy of AI models in complex medical diagnostic tasks.
• This work advances the field of domain adaptation by explicitly handling class shift in multi-source scenarios.