Disentangle domain features for cross-modality cardiac image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) aligns data distributions for models trained on labeled source data to perform on unlabeled target data.
• Existing UDA methods often overlook domain-specific features (DSFs), which can provide valuable complementary information.

Purpose of the Study:

• To propose a new UDA framework for cross-modality image segmentation that leverages both domain-invariant features (DIFs) and DSFs.
• To enhance feature representation and improve segmentation performance on target domains with limited or no labels.

Main Methods:

• The framework disentangles each domain into DIFs and DSFs.
• Self-attention modules enhance DIF representation, while a zero loss minimizes target (source) DSFs in source (target) images.
• Iterative encoding/decoding maintains anatomical consistency, and adversarial learning with discriminators improves image and segmentation quality.

Main Results:

• The proposed UDA framework was validated for cross-modality cardiac segmentation.
• The method demonstrated promising performance and favorable comparison to state-of-the-art approaches in segmentation accuracy.

Conclusions:

• The novel UDA framework effectively utilizes both DIFs and DSFs for improved cross-modality image segmentation.
• The approach shows significant potential for medical imaging applications where labeled data across different modalities is scarce.