Multi-Scale Feature Alignment for Continual Learning of Unlabeled Domains

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) is vital for deep learning in medical fields like histopathology due to limited labeled data.
• Existing UDA methods often focus on single target domains, which is insufficient for long-term applications with evolving data.

Purpose of the Study:

• To develop a continual unsupervised domain adaptation method capable of sequential adaptation to multiple target domains.
• To address the challenge of adapting models without storing previously seen data, adhering to data protection regulations.

Main Methods:

• Proposed a novel approach using generative feature-driven image replay.
• Employed a dual-purpose discriminator for realistic feature generation and promoting feature alignment.
• Evaluated the method on a sequence of three histopathological datasets for tissue-type classification.

Main Results:

• Achieved state-of-the-art results on sequential histopathological tissue-type classification tasks.
• Demonstrated the effectiveness of generative replay and the dual discriminator through extensive ablation studies.
• Showcased a use-case in unsupervised patch-based segmentation of high-resolution tissue images.

Conclusions:

• The proposed continual UDA method effectively handles sequential adaptation to multiple domains without data storage.
• The generative replay and dual discriminator strategy offers a robust solution for UDA in privacy-sensitive medical imaging applications.
• This work advances the application of deep learning in histopathology by enabling continuous model improvement on new datasets.