Multi-parametric artificial neural network fitting of phase-cycled balanced steady-state free precession data

Area of Science:

• Magnetic Resonance Imaging
• Quantitative MRI
• Artificial Intelligence in Medical Imaging

Background:

• Standard T1 and T2 relaxation time quantification using balanced steady-state free precession (bSSFP) is prone to underestimation due to intra-voxel frequency variations.
• Existing methods like MIRACLE can suffer from off-resonance artifacts, limiting accuracy.

Purpose of the Study:

• To develop and validate an artificial neural network (ANN) fitting approach for accurate simultaneous quantification of T1, T2, and field maps (B0, ΔB0) from bSSFP data.
• To assess the ANN's performance with reduced numbers of phase-cycles for accelerated acquisition.

Main Methods:

• A feedforward ANN was trained using whole-brain bSSFP data (3T) with 12, 6, and 4 phase-cycles.
• The ANN input consisted of the Fourier transformed complex bSSFP frequency response (magnitude and phase).
• The target output for training was the multi-parametric set [T1, T2, B0, ΔB0]. ANN predictions were validated against reference values and MIRACLE.

Main Results:

• ANN predictions for T1 and T2 showed excellent agreement with reference values, even with only 4 phase-cycles.
• MIRACLE-based relaxometry exhibited significant off-resonance artifacts with 4 phase-cycles.
• ANN-derived B0 and ΔB0 maps demonstrated high agreement with reference measurements across all tested phase-cycling schemes.

Conclusions:

• ANNs offer a promising method for accurate brain tissue T1 and T2 quantification and reliable field map estimation from bSSFP.
• The bSSFP acquisition can be accelerated by reducing phase-cycles without compromising the robustness of quantitative parameter estimation.