Cross-domain additive learning of new knowledge rather than replacement.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Source-free domain adaptation is crucial in medical imaging due to data privacy and accessibility issues.
• Existing methods often suffer from catastrophic forgetting, losing source domain knowledge during adaptation.
• This limits the robustness and generalizability of models in real-world clinical applications.

Purpose of the Study:

• To develop an additive source-free domain adaptation framework for medical image segmentation.
• To address the challenge of adapting pre-trained models to new domains without access to source data.
• To ensure models retain source domain knowledge while learning from target domain data.

Main Methods:

• A two-stage additive source-free adaptation framework is proposed.
• Domain-invariant features are generalized by constraining pathological structure and semantic consistency.
• Monte-Carlo uncertainty estimation is used to identify and filter segmentation errors.

Main Results:

• The proposed method effectively solves domain offset problems in medical image segmentation.
• Experimental results demonstrate that the model retains its performance on the source domain after adaptation.
• The method shows significant improvements on cross-device polyp and cross-modal brain tumor segmentation datasets.

Conclusions:

• The framework enables additive learning on target and source domains without source data.
• It offers a novel approach for domain adaptation in medical image segmentation.
• This research provides valuable insights for improving model adaptability and robustness in clinical settings.