Unsupervised Domain Adaptation With Variational Approximation for Cardiac Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation is crucial for medical image segmentation when target domain ground truths are unavailable.
• Existing methods often map source and target domain images to a common latent space, reducing discrepancies via adversarial training or metric minimization.
• These approaches face challenges in effectively bridging the distribution gap between different imaging modalities or sequences.

Purpose of the Study:

• To propose a novel unsupervised domain adaptation framework for medical image segmentation.
• To drive latent features of both source and target domains towards a common, parameterized variational Gaussian distribution.
• To enable unsupervised training of segmentation models for target domains with limited or no ground truth data.

Main Methods:

• The framework utilizes two variational auto-encoders (VAEs), one for each domain, incorporating a segmentation module.
• A novel explicit regularization term is introduced to guide the variational approximation and narrow the distribution gap between domains.
• The source segmentation is trained supervisedly, while the target segmentation is trained unsupervisedly within the VAE framework.

Main Results:

• The proposed method demonstrated superior accuracy in cross-modality (CT/MR) and cross-sequence cardiac MR segmentation tasks compared to state-of-the-art approaches.
• The explicit regularization effectively reduced the distribution gap between imaging domains, proving beneficial for unsupervised domain adaptation.
• The framework showed significant potential for improving cardiac segmentation in scenarios lacking target domain labels.

Conclusions:

• The proposed VAE-based unsupervised domain adaptation framework offers a promising solution for medical image segmentation challenges.
• The explicit regularization technique is effective and efficient in aligning feature distributions across domains.
• This approach holds potential for broader applications in medical image analysis where labeled data is scarce.