A confidence-guided Unsupervised domain adaptation network with pseudo-labeling and deformable CNN-transformer for...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) shows promise in medical image segmentation.
• Significant domain differences hinder current UDA methods, necessitating robust cross-domain solutions.
• Existing UDA techniques use fixed feature alignment, failing to adapt to shifting distributions during training.

Purpose of the Study:

• To introduce a novel confidence-guided unsupervised domain adaptation network (CUDA-Net).
• To overcome domain gaps and adapt to shifting feature distributions in medical image segmentation.
• To enhance boundary delineation and detail retention in the target domain.

Main Methods:

• CUDA-Net employs adaptive feature alignment, transitioning from adversarial to pseudo-label-driven alignment.
• A confidence-weighted mechanism refines pseudo-labels, prioritizing high-confidence regions.
• The method tracks cross-domain distribution shifts throughout the training process.

Main Results:

• CUDA-Net demonstrates superior performance on MMWHS17, BraTS2021, and VS-Seg datasets.
• The network outperforms eight leading methods in segmentation accuracy (Dice) and boundary precision (ASD).
• Adaptive alignment and confidence-weighted refinement enhance label reliability and model stability.

Conclusions:

• CUDA-Net offers an efficient and reliable solution for cross-domain medical image segmentation.
• The adaptive approach effectively addresses challenges posed by domain shifts.
• Improved segmentation accuracy and boundary delineation are achieved in target domains.