Contour Transformer Network for One-Shot Segmentation of Anatomical Structures

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of anatomical structures is crucial for medical image analysis.
• Supervised learning methods achieve high accuracy but require extensive expert annotations, posing a scalability challenge.
• Annotation-efficient methods are highly desirable for practical applications.

Purpose of the Study:

• To develop an annotation-efficient, one-shot learning method for anatomical structure segmentation.
• To introduce the Contour Transformer Network (CTN) with an integrated human-in-the-loop mechanism.
• To reduce the dependency on large, expert-labeled datasets in medical imaging.

Main Methods:

• Formulated anatomy segmentation as a contour evolution process.
• Modeled contour evolution using graph convolutional networks (GCNs).
• Developed CTN requiring only one labeled image exemplar and leveraging unlabeled data via novel loss functions for contour consistency.

Main Results:

• CTN significantly outperformed non-learning-based methods in segmenting four different anatomies.
• The one-shot learning approach performed competitively with state-of-the-art fully supervised deep learning methods.
• Minimal human-in-the-loop editing further improved CTN's performance, surpassing fully supervised methods.

Conclusions:

• CTN offers an effective one-shot learning solution for anatomical segmentation, drastically reducing annotation requirements.
• The method demonstrates strong performance and competitiveness with existing supervised techniques.
• The integrated human-in-the-loop mechanism allows for further performance enhancement, highlighting its practical utility.