Entropy-guided partial annotation for cross-domain rib segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for medical imaging, specifically rib segmentation, faces challenges due to scarce expert-annotated Computed Tomography (CT) data.
• Domain shift limits model applicability across different clinical datasets, necessitating costly re-annotation for adaptation.
• Existing semi-supervised methods often involve sample-level annotation, leading to redundancy and high costs for full CT scans.

Purpose of the Study:

• To propose a novel semi-supervised method, Entropy-Guided Partial Annotation (EGPA), to reduce annotation workload for rib segmentation.
• To enhance model training from scratch and cross-domain adaptation by intelligently selecting informative regions for annotation.
• To demonstrate EGPA's effectiveness in achieving high segmentation accuracy with significantly reduced annotation effort.

Main Methods:

• Developed the Entropy-Guided Partial Annotation (EGPA) method, utilizing entropy metrics to identify optimal regions for annotation.
• Integrated contrastive learning, active learning, and self-training strategies within the EGPA framework.
• Evaluated EGPA on both public (RibSegV2) and private chest CT datasets for rib segmentation.

Main Results:

• EGPA achieved Dice scores of 89.5% on the source domain and 90.7% on the target domain.
• Performance closely matched fully supervised models (89.9% and 91.2%) using only 19% and 18% of the annotation workload, respectively.
• Demonstrated significant reduction in annotation cost and time for both initial training and cross-domain adaptation.

Conclusions:

• EGPA substantially reduces annotation effort for rib segmentation, making it more feasible for clinical applications.
• The method effectively addresses domain shift challenges, enabling easier deployment of segmentation tools across different datasets.
• EGPA facilitates the creation of high-quality annotated datasets, promoting standardized and efficient rib analysis in clinical settings.