Comparison and calibration of MP2RAGE quantitative T1 values to multi-TI inversion recovery T1 values

Adam M. Saunders, Michael E. Kim, Chenyu Gao, Lucas W. Remedios, Aravind R. Krishnan, Kurt G. Schilling, Kristin P. O’Grady, Seth A. Smith, Bennett A. Landman

TL;DR

While MP2RAGE is the go-to protocol for fast, high-resolution quantitative T1 mapping in the brain, it suffers from tissue-dependent biases compared to slower "gold standard" methods. This paper introduces a deep learning-based calibration approach using a patch-based ResNet-18 to align MP2RAGE values with multi-TI inversion recovery (IR) values, significantly reducing measurement error across white and gray matter.

What Problem Does This Paper Address?

In the world of MRI, we often deal with "weighted" images (like T1-weighted), where voxel brightness is a qualitative mix of tissue properties and scanner settings. Quantitative MRI, however, aims to measure absolute physical parameters—like the longitudinal relaxation time ($T_1$)—to provide a reproducible, standardized metric of brain health.

The challenge lies in the trade-off between speed and accuracy. The "gold standard" for $T_1$ mapping is multi-TI inversion recovery (IR), which samples the signal recovery curve at many points. It’s highly accurate but painfully slow, often taking several minutes just for a few slices. The alternative, MP2RAGE, is much faster and covers the whole brain, but it relies on a simplified mathematical model (monoexponential) that ignores complex effects like magnetization transfer. This leads to systematic biases: MP2RAGE often underestimates $T_1$ values compared to the gold standard.

Methods

The researchers proposed a two-step pipeline to bridge this gap:

  1. MAP Estimation: Instead of the standard point-estimate for $T_1$, they used a Maximum A Posteriori (MAP) approach. By running Monte Carlo simulations with added Gaussian noise, they generated a posterior distribution for $T_1$. This not only provides the $T_1$ value but also an uncertainty estimate (standard deviation). They also tested MP3RAGE (adding a third readout) to see if more data improved the fit.
  2. Deep Learning Calibration: To fix the systematic bias, they trained a patch-based ResNet-18. The network takes a $5 \times 5$ patch of biased MP2RAGE $T_1$ estimates and learns to output the "corrected" $T_1$ value that matches the multi-TI IR ground truth. This allows the model to use local spatial context to inform the correction.

Datasets

The study used a small but high-quality dataset:

  • Subjects: 4 human subjects.
  • Hardware: 7T Philips Achieva scanner.
  • Data types: Paired MP2RAGE, MP3RAGE, and multi-TI IR (Selective Inversion Recovery) scans, along with $B_1^+$ maps for field inhomogeneity correction.
  • Evaluation: A leave-one-out cross-validation strategy (training on two subjects, validating on one, testing on the final one across four folds).

Key Insights and Ideas

  • The Bias is Tissue-Specific: The error isn't a simple offset. MP2RAGE underestimates $T_1$ by different amounts in white matter (~17%) versus gray matter (~5-11%). This confirms that a "one size fits all" correction factor won't work; you need a method that understands tissue context.
  • Spatial Context Wins: The researchers found that the ResNet performed best with $5 \times 5$ patches. Voxel-wise correction (looking at just one pixel at a time) was less effective, suggesting that the relationship between the two imaging protocols is influenced by the surrounding tissue structure.
  • Uncertainty is a Bonus, Not a Requirement: While the MAP approach provided cool uncertainty maps, adding this uncertainty as an extra channel to the ResNet didn't significantly improve the calibration accuracy. The $T_1$ values themselves contained most of the necessary information for the network.
  • MP3RAGE Trade-offs: Adding a third readout (MP3RAGE) didn't necessarily help. In fact, it showed higher sensitivity to $B_1^+$ inhomogeneities, making the standard MP2RAGE a more robust starting point for calibration.

Key Results

The ResNet-18 calibration dramatically reduced the Root Mean Square Error (RMSE) between the fast MP2RAGE and the gold-standard IR values:

Tissue Type Original RMSE (seconds) Calibrated RMSE (seconds)
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) $0.36 \pm 0.02$ $0.17 \pm 0.03$

Qualitatively, the calibrated maps showed much better agreement with the "ground truth" IR maps, effectively removing the systematic underestimation that has plagued MP2RAGE studies.

Limitations and Future Work

  • Sample Size: With only 4 subjects, the study is a proof-of-concept. While the results are statistically significant, a larger cohort is needed to ensure the model generalizes across different ages and pathologies.
  • Generalization: The model was trained on 7T data from a specific scanner. It is unlikely to work "out of the box" for 3T scanners or different manufacturers (e.g., Siemens vs. GE) without retraining on paired data from those systems.
  • Resolution Loss: There was some loss of image detail during the registration process because the multi-TI IR ground truth has a lower resolution (2mm slices) than the high-res MP2RAGE (0.7mm). Future work could explore super-resolution techniques to maintain the high-res benefits of MP2RAGE while applying the calibration.

Summary updated

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 B<sub>1</sub><sup>+</sup> 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 s to 0.11 ± 0.02 s, subcortical gray matter: from 0.26 ± 0.02 s to 0.10 ± 0.02 s, cortical gray matter: from 0.36 ± 0.02 s to 0.17 ± 0.03 s). 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.
BibTeX
@inproceedings{saunders2025comparison,
  author = {Adam M. Saunders and Michael E. Kim and Chenyu Gao and Lucas W. Remedios and Aravind R. Krishnan and Kurt G. Schilling and Kristin P. O’Grady and Seth A. Smith and Bennett A. Landman},
  title = {{Comparison and calibration of MP2RAGE quantitative T1 values to multi-TI inversion recovery T1 values}},
  booktitle = {Magnetic},
  year = {2025},
  doi = {10.1016/j.mri.2025.110322},
}
Figure