SELF-SEMANTIC CONTOUR ADAPTATION FOR CROSS MODALITY BRAIN TUMOR SEGMENTATION.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) is challenging for semantic segmentation across disparate domains.
• Bridging significant domain gaps requires novel approaches for high-level semantic alignment.

Purpose of the Study:

• To develop an effective UDA framework for semantic segmentation by leveraging low-level edge information.
• To improve cross-modality segmentation of brain tumors using magnetic resonance imaging (MRI).

Main Methods:

• Proposing a multi-task framework integrating contouring and semantic segmentation adaptation networks.
• Utilizing both MRI slices and initial edge maps as input for joint training.
• Employing feature and edge map level adversarial learning for cross-domain alignment.
• Incorporating self-entropy minimization to enhance segmentation performance.

Main Results:

• The proposed framework demonstrates validity and superiority in cross-modality brain tumor segmentation on the BraTS2018 database.
• Leveraging edge information effectively reduces the cross-domain gap compared to direct semantic segmentation.
• The multi-task approach facilitates spatial information guidance for semantic adaptation.

Conclusions:

• Exploiting low-level edge information is a viable precursor task for challenging UDA in semantic segmentation.
• The developed framework offers a robust solution for cross-modality medical image segmentation.
• This method shows significant potential for improving automated analysis of medical scans.