Attentive Continuous Generative Self-training for Unsupervised Domain Adaptive Medical Image Translation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Unsupervised Domain Adaptation (UDA) is crucial for applying models to new data domains.
• Self-training methods excel in discriminative tasks but are under-explored for generative tasks like image translation.
• Domain shift poses a significant challenge in medical imaging and other fields.

Approach:

• Developed a Generative Self-Training (GST) framework for domain adaptive image translation.
• Quantified aleatoric and epistemic uncertainties using variational Bayes for reliable data synthesis.
• Implemented a self-attention mechanism to de-emphasize background regions and an alternating optimization scheme.

Key Points:

• GST enables reliable pseudo-label filtering for generative tasks, unlike traditional self-training.
• Uncertainty quantification ensures the quality of synthesized data for adaptation.
• Self-attention and targeted optimization improve focus on relevant image regions.

Conclusions:

• GST framework significantly improves image translation performance in cross-scanner/center scenarios.
• Outperforms adversarial training UDA methods on tasks like MR image translation.
• Demonstrates the potential of self-training for complex generative domain adaptation problems.