Entropy-driven Adversarial Training For Source-free Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Traditional Unsupervised Domain Adaptation (UDA) methods require source data, limiting real-world applications due to privacy constraints (e.g., healthcare).
• Domain shifts between datasets hinder model generalization and performance in sensitive areas.

Purpose of the Study:

• To propose a novel two-step adversarial Source-Free Unsupervised Domain Adaptation (SFUDA) framework.
• To enable domain adaptation while protecting sensitive source data and mitigating domain shifts.

Main Methods:

• Dividing target data into confident/unconfident samples using prediction entropy and Gumbel softmax.
• Employing a two-step adversarial approach with min-max loss and consistency loss.
• Utilizing a weighted L²-SP regularizer to preserve source domain knowledge.

Main Results:

• The proposed SFUDA framework consistently outperforms other SFUDA methods on domain transfer tasks.
• Achieved competitive results compared to state-of-the-art UDA methods that use source data.
• Demonstrated effectiveness in sensitive applications requiring data privacy.

Conclusions:

• The novel SFUDA framework successfully addresses limitations of traditional UDA methods.
• The approach offers a privacy-preserving solution for domain adaptation in real-world scenarios.
• This framework shows significant potential for sensitive applications like healthcare.