Toward Better Generalization Using Synthetic Data: A Domain Adaptation Framework for T₂ Mapping via Multiple...

Area of Science:

• Medical Imaging
• Machine Learning
• Physics

Background:

• Deep learning for quantitative magnetic resonance imaging (qMRI) often requires extensive real-world training data, which can be scarce.
• Physics-based synthetic data generation for qMRI faces domain gaps, limiting model generalization in real-world applications.
• The multiple overlapping-echo detachment imaging (MOLED) technique allows for rapid single-shot qMRI and transverse relaxation time (T2) quantification.

Purpose of the Study:

• To develop a T2 mapping method for MOLED using domain adaptation to overcome the limitations of synthetic training data.
• To achieve accurate T2 quantification without relying on real-world labeled training samples.
• To reduce the research and development costs associated with qMRI sequence optimization.

Main Methods:

• A domain adaptation strategy was employed for T2 mapping using the MOLED sequence.
• Physics-based simulation was used to generate synthetic data, but the core innovation lies in adapting this data to real-world distributions.
• The proposed method leverages the MOLED sequence's capability for rapid T2 quantification.

Main Results:

• The domain adaptation approach successfully improved the generalization performance of the qMRI model.
• Accurate T2 mapping was achieved without the need for real-world, labeled training data.
• Experimental results showed superior restoration of MR anatomical structures compared to previous methods.

Conclusions:

• Domain adaptation is a viable strategy to bridge the gap between synthetic and real-world data in deep learning for qMRI.
• The proposed MOLED-based T2 mapping method offers an efficient and accurate solution for quantitative imaging.
• This approach enhances the practical applicability of deep learning in rapid qMRI by reducing data acquisition and annotation burdens.