Abstract
While typical qualitative T1-weighted magnetic resonance images reflect scanner and protocol differences, quantitative T1 mapping aims to measure T1 independent of these effects. Changes in T1 in the brain reflect structural changes in brain tissue. Magnetization-prepared two rapid acquisition gradient echo (MP2RAGE) is an acquisition protocol that allows for efficient T1 mapping with a much lower scan time per slab compared to multi-TI inversion recovery (IR) protocols. We collect and register B1-corrected MP2RAGE acquisitions with an additional inversion time (MP3RAGE) alongside multi-TI selective inversion recovery acquisitions for four subjects. We use a maximum a posteriori (MAP) T1 estimation method for both MP2RAGE and compare to typical point estimate MP2RAGE T1 mapping, finding no bias from MAP MP2RAGE but a sensitivity to B1 inhomogeneities with MAP MP3RAGE. We demonstrate a tissue-dependent bias between MAP MP2RAGE T1 estimates and the multi-TI inversion recovery T1 values. To correct this bias, we train a patch-based ResNet-18 to calibrate the MAP MP2RAGE T1 estimates to the multi-TI IR T1 values. Across four folds, our network reduces the RMSE significantly (white matter: from 0.30 +/- 0.01 seconds to 0.11 +/- 0.02 seconds, subcortical gray matter: from 0.26 +/- 0.02 seconds to 0.10 +/- 0.02 seconds, cortical gray matter: from 0.36 +/- 0.02 seconds to 0.17 +/- 0.03 seconds). Using limited paired training data from both sequences, we can reduce the error between quantitative imaging methods and calibrate to one of the protocols with a neural network.
Introduction
In clinical and research MRI, voxel brightness in standard "weighted" images is a qualitative mixture of tissue properties and scanner-specific settings. Quantitative MRI (qMRI) seeks to overcome this by measuring absolute physical parameters, such as the longitudinal relaxation time ($T_1$), providing a standardized metric for brain health.
The primary challenge in qMRI is the trade-off between acquisition speed and measurement accuracy. The "gold standard" for $T_1$ mapping is multi-TI inversion recovery (IR), which samples the signal recovery curve at multiple points. While highly accurate, it is prohibitively slow for whole-brain coverage. Magnetization-prepared two rapid acquisition gradient echo (MP2RAGE) offers a faster, high-resolution alternative but often suffers from systematic biases due to simplified mathematical models that ignore complex tissue effects like magnetization transfer. This study addresses these biases by introducing a deep learning-based calibration approach.
Methods
Data Acquisition and MAP Estimation
The study utilized a 7T Philips Achieva scanner to collect paired MP2RAGE, MP3RAGE (adding a third readout), and multi-TI IR scans from four subjects, alongside $B_1^+$ maps for field inhomogeneity correction.
Instead of standard point estimates, a Maximum A Posteriori (MAP) approach was employed for $T_1$ estimation. By performing Monte Carlo simulations with Gaussian noise, the researchers generated posterior distributions for $T_1$, providing both a mean value and an uncertainty estimate (standard deviation).
Deep Learning Calibration
To correct the systematic underestimation in MP2RAGE, a patch-based ResNet-18 was trained. The network processed $5 \times 5$ patches of biased MP2RAGE $T_1$ estimates to predict "corrected" values matching the multi-TI IR ground truth. This spatial context allowed the model to account for local tissue structures. Evaluation was performed using a leave-one-out cross-validation strategy across four folds.
Results
Tissue-Specific Bias
The analysis confirmed that MP2RAGE bias is tissue-dependent rather than a simple global offset. MP2RAGE was found to underestimate $T_1$ by approximately 17% in white matter and 5-11% in gray matter. Furthermore, while MP3RAGE provided more data points, it exhibited higher sensitivity to $B_1^+$ inhomogeneities, making calibrated MP2RAGE a more robust choice.
Calibration Performance
The ResNet-18 calibration significantly reduced the Root Mean Square Error (RMSE) across all examined tissue types:
| Tissue Type | Original RMSE (s) | Calibrated RMSE (s) |
|---|---|---|
| White Matter (WM) | $0.30 \pm 0.01$ | $0.11 \pm 0.02$ |
| Subcortical Gray Matter (SGM) | $0.26 \pm 0.02$ | $0.10 \pm 0.02$ |
| Cortical Gray Matter (CGM) | $0.36 \pm 0.02$ | $0.17 \pm 0.03$ |
Spatial context proved critical; voxel-wise correction was less effective than the $5 \times 5$ patch-based approach. Interestingly, while the MAP method provided uncertainty maps, including them as input to the ResNet did not significantly improve calibration accuracy beyond using the $T_1$ values alone.
Discussion and Conclusion
This work demonstrates that deep learning can effectively calibrate fast qMRI protocols to match slower gold-standard measurements. By using a patch-based ResNet-18, the systematic, tissue-dependent biases of MP2RAGE were substantially mitigated.
Limitations include the small sample size (N=4) and the specific focus on 7T hardware. Future research should investigate the generalizability of this calibration across different field strengths (e.g., 3T) and scanner manufacturers. Additionally, addressing the resolution mismatch between high-res MP2RAGE and lower-res IR ground truth through super-resolution techniques could further refine the calibration process.
Summary updated