Uncertainty-aware domain alignment for anatomical structure segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of anatomical structures in medical images is vital for disease detection.
• Deep neural networks often face performance degradation due to domain shift across different medical imaging modalities or devices.
• Unsupervised Domain Adaptation (UDA) aims to bridge this gap but requires robust methods.

Purpose of the Study:

• To propose an uncertainty-aware domain alignment framework to address the domain shift problem in cross-domain UDA for medical image segmentation.
• To enhance segmentation performance by effectively utilizing uncertainty information.
• To develop strategies for optimal target sample selection in UDA.

Main Methods:

• Designed an Uncertainty Estimation and Segmentation Module (UESM) for uncertainty map estimation.
• Proposed an Uncertainty-aware Cross Entropy (UCE) loss to improve segmentation in uncertain regions.
• Developed an Uncertainty-aware Self-Training (UST) strategy for guided target sample selection.
• Implemented an Uncertainty Feature Recalibration Module (UFRM) to minimize cross-domain discrepancies.

Main Results:

• The proposed UESM demonstrated efficiency and effectiveness in uncertainty estimation for UDA.
• Achieved state-of-the-art performance on both cross-modality (MMWHS 2017 cardiac dataset) and cross-device (private OCT dataset) benchmarks.
• The uncertainty-aware approach significantly boosted segmentation accuracy in challenging domain shift scenarios.

Conclusions:

• The developed uncertainty-aware domain alignment framework successfully mitigates the domain shift problem in medical image segmentation.
• The UESM, UCE loss, and UST strategy collectively contribute to superior performance in UDA tasks.
• This approach offers a promising direction for robust and accurate medical image analysis across diverse data sources.