Dual Discriminator-Based Unsupervised Domain Adaptation Using Adversarial Learning for Liver Segmentation on...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Multiphase computed tomography (CT) is crucial for liver disease diagnosis.
• Manual segmentation of multiphase CT images is time-consuming and labor-intensive due to domain differences across phases.
• Accurate segmentation is vital for diagnosis and surgical planning.

Purpose of the Study:

• To develop an unsupervised domain adaptation method for accurate liver segmentation on multiphase CT images without annotations.
• To improve the efficiency and reduce the cost of liver segmentation in medical imaging.

Main Methods:

• Proposed a dual discriminator-based unsupervised domain adaptation (DD-UDA) framework.
• Framework includes a task-specific generator and two discriminators.
• Domain adaptation performed at both feature and output levels to minimize domain distribution differences.

Main Results:

• Demonstrated the effectiveness of the DD-UDA method using public (PV phase) and private multiphase CT data.
• Achieved significant improvements in segmentation accuracy across all CT image phases compared to methods without unsupervised domain adaptation (5%, 8%, 6% increase).

Conclusions:

• The DD-UDA method enables efficient and accurate liver segmentation on multiphase CT images.
• This approach addresses the challenge of domain shift in multiphase CT imaging.
• The findings support the clinical relevance of automated segmentation for diagnosis and surgical support.