B₁ mapping using pre-learned subspaces for quantitative brain imaging.

Area of Science:

• Medical Imaging
• Machine Learning
• Quantitative Brain Imaging

Background:

• B1 field inhomogeneity is a significant challenge in quantitative brain imaging, affecting accuracy.
• Accurate estimation of both transmitter (B1t) and receiver (B1r) fields is crucial for effective B1 inhomogeneity correction.

Purpose of the Study:

• To develop a novel machine learning-based method for estimating transmitter and receiver B1 fields.
• To improve the correction of B1 inhomogeneity effects in quantitative brain imaging.

Main Methods:

• A subspace model-based machine learning approach was employed to estimate B1t and B1r fields.
• Probabilistic subspace models captured scan-dependent B1 field variations, learning from pre-scanned data.
• New experimental data B1 field estimation was performed via linear optimization with prior constraints.

Main Results:

• The method generated high-quality B1 maps in phantoms and healthy subjects.
• B1 correction significantly improved T1 and proton density (PD) maps from SPGR data.
• In brain tumor patients, the method demonstrated more accurate and robust B1 estimation and correction than conventional techniques.
• Application to MRSI data from tumor patients yielded improved neurometabolite maps with better tissue differentiation.

Conclusions:

• A novel machine learning method using probabilistic subspace models was developed for B1 field estimation.
• This approach offers a more robust solution for correcting B1 inhomogeneity in practical quantitative brain imaging applications.