Abstract
Magnetic Resonance Fingerprinting (MRF) is a rapid quantitative imaging technique sensitive to $B_1^+$ and $B_0$ field inhomogeneities, which can bias tissue parameter estimation. We present the Sequence Adaptive Field-imperfection Estimation (SAFE) framework, a deep learning approach that retrospectively estimates these field maps directly from MRF data. By utilizing a simulation-based training pipeline and subspace coefficient maps, SAFE achieves sequence-agnostic performance, significantly reducing parameter bias and artifacts without requiring additional calibration scans.
Introduction
Magnetic Resonance Fingerprinting (MRF) enables simultaneous mapping of multiple tissue properties, such as $T_1$ and $T_2$ relaxation times, in a single acquisition. However, the accuracy of MRF is often compromised by field imperfections, specifically inhomogeneities in the static magnetic field ($B_0$) and the radiofrequency excitation field ($B_1^+$). These imperfections lead to blurring and significant biases in quantitative maps. While prospective calibration scans can mitigate these effects, they increase scan time and cannot be applied retrospectively to existing datasets.
Methods
The SAFE framework employs a U-Net-inspired convolutional neural network (CNN) to estimate field maps from MRF data. A key innovation is the use of subspace coefficient maps as inputs. These maps represent a compressed version of the complex signal fingerprints, capturing essential physics while reducing computational load.
Sequence-Agnostic Training
To ensure generalizability across different pulse sequences, SAFE uses a simulation-based training strategy. High-quality quantitative maps ($T_1$, $T_2$, PD, $B_1^+$, $B_0$) are used to simulate MRF signals for specific sequence parameters (e.g., flip angles, TR). This allows the model to be adapted to new sequences by simply re-running the simulation pipeline rather than requiring new ground-truth acquisitions.
Physics-Informed Modeling and Auxiliary Tasks
The simulation incorporates $B_1^+$ effects via dictionary models and $B_0$ effects through spatial blurring. During training, the network performs auxiliary tasks, generating corrected tissue parameter maps alongside the field maps. This multi-task learning approach leverages the spatial smoothness of field maps to improve estimation accuracy.
Results
Validation on healthy volunteers and clinical cohorts demonstrated that SAFE's predicted $B_1^+$ and $B_0$ maps closely match ground-truth calibration scans. In regions with significant $B_1^+$ inhomogeneity, SAFE reduced $T_2$ estimation bias by 24.2%. Qualitatively, the framework effectively mitigated $B_0$-induced blurring in challenging anatomical regions like the temporal lobes and optic nerves. Crucially, the model successfully generalized across three distinct MRF sequences without sequence-specific training data.
Discussion and Conclusion
SAFE provides a robust, retrospective solution for field imperfection correction in MRF. By operating on subspace coefficients and utilizing a simulation-driven training approach, it bypasses the need for sequence-specific datasets and additional calibration scans. While currently focused on brain imaging at 3T, future work could extend this framework to other organs and system imperfections like eddy currents. SAFE represents a significant step toward making quantitative MRI more reliable and accessible for retrospective analysis of large-scale clinical studies.
Summary updated