Generative appearance replay for continual unsupervised domain adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels with large labeled datasets, but real-world data presents challenges like incremental availability, domain shifts, and lack of labels.
• Medical applications face strict privacy regulations prohibiting the retention of previously seen data.
• Unsupervised domain adaptation (UDA) is crucial for leveraging unlabeled data across different domains.

Purpose of the Study:

• To address challenges in unsupervised segmentation within continual learning scenarios involving domain shift.
• To develop a method that adapts segmentation models sequentially to new domains using unlabeled data.
• To enable adaptation without requiring access to previously seen data, enhancing practical applicability.

Main Methods:

• Introduction of GarDA (Generative Appearance Replay for continual Domain Adaptation), a generative-replay based approach.
• Sequential adaptation of a segmentation model to new domains with unlabeled data.
• Leveraging and consolidating information from multiple domains through continual adaptation.

Main Results:

• GarDA substantially outperforms existing techniques in unsupervised segmentation across multiple domains.
• The method demonstrates effectiveness in continual learning scenarios with domain shift.
• Successful evaluation on three datasets involving different organs and modalities.

Conclusions:

• GarDA offers a practical solution for unsupervised segmentation in privacy-constrained, continual learning settings.
• The generative-replay approach effectively handles domain shift without data retention.
• This work advances the field of incremental UDA by removing the need for prior data access.