Generative Self-training for Cross-domain Unsupervised Tagged-to-Cine MRI Synthesis.

Area of Science:

• Deep Learning
• Medical Imaging
• Computer Vision

Background:

• Unsupervised domain adaptation (UDA) using self-training is effective for discriminative tasks but underexplored for generative tasks like image synthesis.
• Domain shift poses challenges when applying deep learning models trained on source data to unlabeled target domains.

Purpose of the Study:

• To propose a novel Generative Self-Training (GST) Unsupervised Domain Adaptation (UDA) framework for cross-domain image synthesis.
• To address limitations of existing self-training UDA methods in generative applications.

Main Methods:

• Developed a GST UDA framework utilizing continuous value prediction and a regression objective for image synthesis.
• Implemented pseudo-label filtering with an uncertainty mask and employed variational Bayes learning to quantify predictive confidence.
• Utilized a round-based alternative optimization scheme for fast test-time adaptation.

Main Results:

• The proposed GST framework demonstrated significant improvements in synthesis quality for tagged-to-cine magnetic resonance imaging (MRI) synthesis.
• GST outperformed popular adversarial training UDA methods when adapting to new target domains with unlabeled data.
• Validation on MRI data acquired from different scanners/centers confirmed the framework's effectiveness.

Conclusions:

• The novel GST UDA framework offers a robust solution for cross-domain image synthesis, particularly in medical imaging applications.
• GST effectively handles domain shift in generative tasks by leveraging continuous value prediction and uncertainty quantification.
• The proposed method shows strong potential for improving the quality and generalizability of synthesized medical images.