Introduction

T1 mapping allows for quantitative tissue evaluation to assist in the diagnosis of diseases.^1,2 Traditional T1 mapping is performed using a Cartesian MOLLI sequence, which only allows acquisition of 1-2 slices within a 20s breath hold period (BHP). Recently, a radial Look-Locker (LL) technique was proposed which demonstrates improved slice efficiency for acquisition of up to 11 slices within a BHP.³ Unfortunately, this only provides partial coverage of the abdomen. The goal of this work is to enable full abdominal coverage within a single BHP. This is achieved by performing short sampling of the T1 recovery curve (T1RC), followed by convolutional neural network (CNN) T1 estimation.

Methods

Technique: A 2D radial LL T1 mapping technique was implemented (Figure 1). A baseline sequence (Figure 1A) utilizing long sampling of the T1RC is used to acquire reference and training data for the CNN. Here, radial views are divided into 16-view groups to produce $$$N$$$ TI groups ($$$K_1,…,K_N$$$). These are used in a locally low rank (LLR)⁴ reconstruction to obtain the corresponding TI images ($$$I_1,…,I_N$$$). Single pixel fitting (SPF) using a non-linear least squares method is used to generate T1 maps.

The accelerated sequence (Figure 1B) employs shortened sampling of the T1RC to improve slice efficiency. This results in fewer radial views being acquired thus, grouping of views yields $$$\hat{N}<N$$$ TI groups ($$$K_1,…,K_\hat{N}$$$). LLR reconstruction is used to obtain a reduced set of TI images ($$$\hat{I_1},...,\hat{I_\hat{N}}$$$).

To overcome T1 estimation error due to reduced signal quality from the short T1RC, a supervised CNN model is trained to estimate T1 maps directly from the reduced set of TI images. The $$$N$$$ TI groups obtained from the long T1RC data, acquired using the baseline LL sequence, are used to obtain the T1 maps used as the CNN labels. The number of views within the long T1RC dataset is then retrospectively truncated to produce a dataset with $$$\hat{N}<N$$$ TI groups. This simulates data acquired through shortened sampling of the T1RC. The TI images reconstructed from the truncated dataset serve as the CNN input. For testing, we used short T1RC data prospectively acquired with the accelerated LL sequence to obtain $$$\hat{N}$$$ TI images. These are given to the pre-trained network to generate a T1 map estimate.

In vivo imaging: Data were acquired at 3T (MAGNETOM Skyra, Siemens) with TR=3.3ms, TE=1.75ms, $$$\alpha$$$=10° and pixel-resolution=1.56 x 1.56 x 8mm. Training data: Data from 20 subjects were acquired with 21 slices per subject over 3 breath-holds using the baseline LL sequence. T1RC was 2.5s and consisted of 768 radial views. The dataset was retrospectively truncated to 576, 512, 304, 256, 192 and 128 views to simulate datasets at T1RC lengths of 1.9s, 1.69s, 1.0s, 0.84s, 0.63s and 0.42s, respectively. A separate CNN was trained using each dataset. Testing Data: Data from 5 subjects were acquired with the accelerated LL sequence. The T1RC=0.84s (256 radial views) yielded 21 slices within a single 20s BHP. An optimized slice interleaving scheme⁵ was used to reduce T1 saturation. Reference data for each subject were acquired using the training data protocol (T1RC=2.5s).

Results and Discussion

Figure 3A demonstrates the impact of T1RC length on T1 estimation. Here the various T1RC lengths were generated by retrospectively cutting the reference 2.5s curve. We used (i) conventional SPF to fit the TI images and (ii) CNN to estimate T1 maps directly from the network. SPF achieved T1 estimation error below 5% (4.0%) in liver for T1RC=1.69s but exceeds 5% for muscle (6.0%) and spleen (7.4%) even for T1RC=1.9s. On the other hand, CNN T1 estimation error remained below 5% across liver (3.8%), muscle (3.4%) and spleen (4.6%) down to T1RC=0.84s. Qualitative evaluation (Figure 3B) demonstrates the superior quality of CNN generated T1 maps across T1RC length.

Based on the above result, a single BHP accelerated acquisition protocol was developed using T1RC=0.84s combined with optimized slice interleaving to allow acquisition of 21 slices with 8mm slice thickness. This protocol provided full abdominal coverage (168mm) with excellent spatial resolution (1.56x1.56x8mm). Figure 4A demonstrates the consistency in estimated T1 across liver (674 ± 15.9ms), muscle (1115 ± 22.1ms) and spleen (1174 ± 23.4ms) for a single subject, indicating no T1 saturation due to the advanced slice interleaving scheme used. Further validation can be seen by qualitative evaluation of the T1 maps obtained from the acquired slices (Figure 4B).

Figure 5 shows correlation and Bland-Altman analysis for SPF and CNN T1 estimates for data acquired on 5 subjects using the single BHP protocol. Systematic over estimation of T1 is observed for SPF, with an average T1 error of 10.2%/20.7%/16.5% across liver/muscle/spleen, respectively, with a coefficient of variation of 11%. The CNN achieved significant improvements, with average T1 error of 0.9%/1.4%/4.3% across liver/muscle/spleen, respectively, and a coefficient of variation of 2.7%.

Conclusions

An accelerated T1 mapping framework was developed which allows for full abdominal coverage within a single BHP. This technique should allow improved efficiency in clinical applications of T1 mapping compared to the current standard.

References

1. Kim JW, Lee YS, Park YS, Kim BH, Lee SY, Yeon JE, Lee CH. Multiparametric MR Index for the Diagnosis of Non-Alcoholic Steatohepatitis in Patients with Non-Alcoholic Fatty Liver Disease. Sci Rep. 2020;10(1):2671.

2. Huang Y, Sadowski EA, Artz NS, et al. Measurement and comparison of T1 relaxation times in native and transplanted kidney cortex and medulla. J Magn Reson Imaging. 2011;33:1241-1247.

3. Goerke U, Ahanonu E, Keerthivasan M, Bilgin A, Deshpande V, Altbach MI, Inversion Recovery Look-Locker T1-Mapping for Abdominal Imaging: How Many Slices Can One Fit in a Single Breath-Hold? Proceedings of Annual Meeting of the ISMRM, 30:107, 2022.

4. https://mrirecon.github.io/bart/

5. Li Z, Bilgin A, Johnson K, Galons JP, Martin DR, and Altbach MI, Rapid high-resolution T1 mapping using highly accelerated radial steady-state free-precession acquisition, J Magn Reson Imaging, 2019 Jan;49(1):239-252. doi: 10.1002/jmri.26170. Epub 2018 Aug 24. PMID: 30142230.