Unsupervised cross-modality domain adaptation via source-domain labels guided contrastive learning for medical image...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) is crucial for improving model performance on target domains using source domain knowledge.
• Medical image segmentation often faces challenges due to domain shifts across different modalities.
• Existing UDA methods struggle to effectively bridge the gap in medical image segmentation tasks.

Purpose of the Study:

• To propose a unified domain adaptation framework for cross-modality medical image segmentation.
• To enhance model robustness and alignment from both image and feature perspectives.
• To leverage source domain information for improved target domain segmentation without labeled target data.

Main Methods:

• Fine-tuning Fourier-based Contrastive Style Augmentation (FCSA) for robust image alignment.
• Designing Source-domain Labels Guided Contrastive Learning (SLGCL) for feature alignment.
• Incorporating a generative adversarial network for spatial and contextual consistency.

Main Results:

• The proposed framework achieves superior performance in cross-modality medical image segmentation.
• Demonstrated significant improvements over state-of-the-art UDA methods on a whole heart segmentation task.
• The novel approach effectively aligns image and feature distributions, enhancing segmentation accuracy.

Conclusions:

• The unified UDA framework offers a powerful solution for cross-modality medical image segmentation.
• The integration of FCSA and SLGCL effectively addresses domain shift challenges.
• This work represents a significant advancement in unsupervised domain adaptation for medical imaging.