Model-based spatiotemporal synthetic data generation framework and deep-learning reconstruction for real-time MRI...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Deep learning for quantitative MRI (qMRI) reconstruction often faces data scarcity.
• Current synthetic data methods primarily focus on 2D reconstruction, neglecting spatiotemporal correlations in dynamic imaging.
• Leveraging spatiotemporal data can significantly improve reconstruction quality in real-time dynamic qMRI.

Purpose of the Study:

• To develop a model-based spatiotemporal synthetic data generation framework for real-time dynamic qMRI reconstruction.
• To enable accurate tracking of dynamic changes in oxygen extraction fraction (OEF) using 3D spatiotemporal reconstruction.
• To enhance the quality and efficiency of qMRI reconstruction by utilizing synthetic data and advanced imaging techniques.

Main Methods:

• Proposed a model-based spatiotemporal synthetic data generation framework for supervised learning-based reconstruction.
• Integrated an ultra-fast multiple overlapping-echo detachment (MOLED) imaging technique.
• Developed a 3D spatiotemporal reconstruction method with constraints for parameter alignment and singular-value subspace consistency.

Main Results:

• Validated the method's accuracy for T2 and T2* mapping in numerical, phantom, and human brain experiments.
• Demonstrated superior performance compared to traditional 2D spatial and 3D spatiotemporal reconstruction methods.
• Successfully enabled dynamic OEF tracking during breath-hold and oxygen inhalation, showing robustness in real-time scenarios.

Conclusions:

• The proposed framework offers an effective and robust solution for real-time qMRI.
• The MOLED sequence facilitates precise, dynamic measurement of brain oxygen metabolism.
• This approach has potential applications in other real-time dynamic quantitative reconstruction tasks and provides insights into cerebral physiology.