FPL+: Filtered Pseudo Label-Based Unsupervised Cross-Modality Adaptation for 3D Medical Image Segmentation

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Adapting medical image segmentation models to new domains is crucial for clinical applications.
• Unsupervised Domain Adaptation (UDA) is valuable due to the high cost of medical image annotation.
• Existing UDA methods struggle with insufficient supervision in the target domain.

Purpose of the Study:

• To propose an enhanced Filtered Pseudo Label (FPL+)-based UDA method for 3D medical image segmentation.
• To improve cross-domain transferability of segmentation models using only unlabeled target domain data.
• To address the limitations of insufficient supervision in current UDA techniques.

Main Methods:

• Utilized cross-domain data augmentation to create a dual-domain training set.
• Employed domain-specific batch normalization layers for domain shift handling and invariant feature learning.
• Generated high-quality pseudo labels for target domain images.
• Implemented image-level weighting (uncertainty estimation) and pixel-level weighting (dual-domain consensus) to refine segmentation using pseudo labels and mitigate noise.

Main Results:

• The proposed FPL+ method demonstrated superior performance compared to ten state-of-the-art UDA methods.
• Achieved results comparable to or even better than fully supervised methods in the target domain across multiple datasets.
• Successfully segmented Vestibular Schwannoma, brain tumors, and whole hearts in multi-modal datasets.

Conclusions:

• The enhanced FPL+ method effectively adapts 3D medical image segmentation models to new domains without labeled target data.
• The proposed weighting strategies significantly improve segmentation accuracy by addressing noisy pseudo labels.
• This approach offers a promising solution for robust and efficient medical image segmentation in diverse clinical settings.