Adapting Off-the-Shelf Source Segmenter for Target Medical Image Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Unsupervised domain adaptation (UDA) transfers knowledge from labeled source to unlabeled target domains.
• Limited access to source data during adaptation poses challenges.
• Existing UDA methods often require source data, limiting practical application.

Purpose of the Study:

• To develop a source-free UDA framework for segmentation tasks.
• To adapt pre-trained segmentation models to unseen target domains without source data.
• To improve the performance and applicability of UDA in data-scarce scenarios.

Main Methods:

• Proposes an adaptive batch-wise normalization statistics adaptation framework.
• Gradually adapts low-order batch statistics (mean, variance) using exponential decay.
• Enforces consistency of high-order batch statistics (scaling, shifting) via optimization.
• Adaptively measures channel transferability to balance contributions.

Main Results:

• Outperforms existing source-relaxed UDA methods in cross-subtype segmentation on BraTS 2018.
• Achieves comparable results to supervised UDA in cross-modality segmentation.
• Demonstrates the framework's effectiveness without requiring source domain data.

Conclusions:

• The proposed source-free UDA framework enables effective model adaptation without source data.
• The method offers a practical solution for UDA in segmentation, especially with data privacy or storage constraints.
• The framework is compatible with other unsupervised learning methods, enhancing its versatility.