SA-Seg: Annotation-Efficient Segmentation for Airway Tree Using Saliency-Based Annotation.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Radiology

Background:

• Airway tree segmentation is crucial for clinical practice but is hindered by complex structures and annotation challenges.
• Existing annotation-efficient methods are often unsuitable for the airway's unique characteristics.

Purpose of the Study:

• To develop an annotation-efficient segmentation method for airway trees that enhances both efficiency and completeness.
• To address the limitations of current methods in handling complex airway structures.

Main Methods:

• A saliency-based annotation approach requiring only high-saliency regions to be annotated.
• A positive-unlabeled learning-inspired probabilistic model with score and bias functions, implemented using convolutional neural networks and an EM algorithm.
• Modeling the dependency between key annotation elements to learn from biased, weak annotations.

Main Results:

• Achieved an 89% reduction in annotation time compared to traditional methods.
• Significantly minimized the performance discrepancy between weak and fully annotated datasets.
• Demonstrated the method's effectiveness in improving airway tree segmentation efficiency and completeness.

Conclusions:

• The proposed method offers a substantial improvement in annotation efficiency for airway tree segmentation.
• The approach shows significant potential for practical clinical applications due to its time-saving and accuracy-enhancing features.