A unified model for reconstruction and R₂^* mapping of accelerated 7T data using the quantitative recurrent inference...

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Quantitative MRI (qMRI) at 7 Tesla enables subcortical structure analysis but requires long scan times.
• Deep learning offers potential to accelerate qMRI by leveraging data redundancy and relaxometry models.

Purpose of the Study:

• To develop and evaluate a deep learning approach, the quantitative Recurrent Inference Machine (qRIM), for accelerated qMRI R2* mapping.
• To assess the impact of network architecture and data undersampling on reconstruction accuracy and biological signal preservation.

Main Methods:

• Proposed qRIM with a unified forward model for joint reconstruction and R2* mapping from sparse, multi-echo gradient echo data.
• Compared qRIM against a quantitative End-to-End Variational Network (qE2EVN), U-Net, and Compressed Sensing.
• Evaluated performance using high-resolution brain data from a lifespan cohort at 7T.

Main Results:

• qRIM significantly reduced R2* reconstruction error from undersampled data compared to sequential methods.
• Reconstruction error decreased with higher acceleration factors, indicating greater benefit for sparser data.
• Subcortical maturation signals were preserved up to an acceleration factor of 9, outperforming other methods in accuracy and bias.

Conclusions:

• The qRIM, integrating a unified forward model, effectively accelerates relaxometry by exploiting data redundancy and shared information.
• This deep learning approach facilitates efficient and accurate quantitative MRI, particularly for subcortical analysis in accelerated protocols.