Source-Free Active Domain Adaptation via Influential-Points-Guided Progressive Teacher for Medical Image...

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Domain adaptation is crucial for medical image segmentation models to generalize across different datasets.
• Source-Free Active Domain Adaptation (SFADA) addresses data scarcity and privacy concerns by selecting limited target samples for annotation without source data access.
• Existing SFADA methods lack robust class-balanced sample selection and effective model adaptation strategies.

Purpose of the Study:

• To develop a novel SFADA framework for medical image segmentation that enhances generalization to new domains.
• To address the limitations of existing SFADA methods in class-balanced sample selection and robust model adaptation.
• To identify and leverage 'influential points'—boundary samples with source-like semantics and high predictive discrepancy—for efficient learning.

Main Methods:

• Proposed a slice-wise framework utilizing influential points learning, detecting source-like samples to preserve source knowledge.
• Introduced an adaptive K-nearest neighbor algorithm for constructing source-like sample neighborhoods for knowledge transfer.
• Developed a class-balanced Kullback-Leibler divergence to rank target samples by influence, enabling selection of diverse, high-ranked influential points for annotation.
• Designed a progressive teacher model with curriculum learning to generate and utilize pseudo-labels, mitigating error accumulation and progressively leveraging reliable supervision.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches in medical image segmentation domain adaptation.
• Effective identification and utilization of influential points led to robust model adaptation.
• The framework achieved superior performance even with a minimal labeling budget (2.5%).

Conclusions:

• The developed SFADA framework effectively addresses challenges in medical image segmentation, particularly with limited annotated data and privacy constraints.
• Identifying and utilizing influential points is a key strategy for efficient and class-balanced domain adaptation.
• The progressive teacher model and curriculum learning enhance the robustness and accuracy of the adaptation process.