One for multiple: Physics-informed synthetic data boosts generalizable deep learning for fast MRI reconstruction.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) offers radiation-free anatomical insights but suffers from long scan times.
• K-space undersampling accelerates MRI but introduces artifacts requiring complex reconstruction.
• Current deep learning (DL) methods for fast MRI reconstruction face challenges with data acquisition costs, privacy, and generalization across diverse scenarios.

Purpose of the Study:

• To introduce a novel framework, Physics-Informed Synthetic data learning Framework (PISF), for generalizable and fast MRI reconstruction.
• To overcome limitations of existing DL approaches by reducing reliance on large, real-world MRI datasets.
• To enable a single DL model for high-quality reconstruction across multiple imaging parameters and centers.

Main Methods:

• Developed a Physics-Informed Synthetic data learning Framework (PISF) for fast MRI reconstruction.
• Decomposed 2D image reconstruction into multiple 1D problems, starting with 1D synthetic data synthesis for enhanced generalization.
• Employed enhanced learning techniques for training DL models on synthetic data.

Main Results:

• Achieved in vivo MRI reconstructions comparable or superior to models trained on real data, reducing real-world data needs by up to 96%.
• Demonstrated remarkable generalizability with a single PISF model across 4 sampling patterns, 5 anatomies, 6 contrasts, 5 vendors, and 7 centers.
• Validated adaptability to neuro and cardiovascular patient populations through expert evaluations.

Conclusions:

• PISF provides a feasible and cost-effective solution for fast MRI reconstruction.
• The framework significantly enhances the generalizability of deep learning in diverse MRI applications.
• PISF facilitates wider adoption of accelerated MRI techniques, improving diagnostic accessibility.