Memory consistent unsupervised off-the-shelf model adaptation for source-relaxed 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.
• Accessing labeled source data is often restricted due to privacy or IP concerns.
• Existing UDA methods typically require joint training on data from both domains.

Purpose of the Study:

• To propose an "off-the-shelf (OS)" UDA (OSUDA) framework for image segmentation.
• To adapt models to a target domain without requiring access to the source domain data.
• To develop a novel batch-wise normalization (BN) statistics adaptation framework for this purpose.

Main Methods:

• A novel batch-wise normalization (BN) statistics adaptation framework is proposed.
• Low-order BN statistics (mean, variance) are adapted using exponential momentum decay.
• High-order BN statistics (scaling, shifting) are enforced for consistency.
• Channel-wise transferability is quantified adaptively.
• Unsupervised self-entropy minimization and memory-consistent self-training are incorporated.

Main Results:

• The OSUDA framework was evaluated on cross-modality/subtype brain tumor segmentation and cardiac MR to CT segmentation.
• The proposed memory-consistent OSUDA outperforms existing source-relaxed UDA methods.
• It achieves performance comparable to UDA methods that utilize source data.

Conclusions:

• The developed OSUDA framework effectively adapts image segmentation models to new domains without source data.
• The method offers a viable solution for UDA scenarios with restricted access to labeled source data.
• This approach demonstrates strong performance, rivaling traditional UDA methods.