Unsupervised Domain Adaptation for Small Bowel Segmentation using Disentangled Representation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate small bowel segmentation is crucial for diagnosing gastrointestinal conditions.
• Unsupervised domain adaptation methods are needed to address variations in CT scan data, such as the presence or absence of oral contrast.
• Existing methods often struggle with adapting to these domain shifts effectively.

Purpose of the Study:

• To develop a novel unsupervised domain adaptation method for small bowel segmentation.
• To improve the controllability and performance of domain adaptation by disentangling feature types.
• To enhance the transferability of features across different abdominal CT scan domains.

Main Methods:

• A unique two-stream auto-encoding architecture was employed to disentangle intensity and non-intensity features.
• Selective adaptation of non-intensity features was performed to leverage more transferable information.
• Segmentation prediction was achieved by aggregating the disentangled features.

Main Results:

• The proposed method demonstrated significant improvements in small bowel segmentation accuracy across different domains.
• Performance gains were observed in terms of three distinct evaluation metrics compared to baseline methods.
• Feature disentanglement proved effective in enhancing domain adaptation for CT image segmentation.

Conclusions:

• The novel feature disentanglement approach offers a more controllable and effective solution for unsupervised domain adaptation in medical imaging.
• This method advances small bowel segmentation, bringing it closer to reliable clinical application.
• The selective adaptation of non-intensity features shows promise for improving generalization in medical image analysis tasks.