Introduction

Modified Look-Locker inversion recovery (MOLLI) T₁ mapping is routinely used in clinical practice for quantitatively assessing the focal and diffuse fibrosis in diseased myocardium (1,2). MOLLI is well known for the T₁ underestimation and requires a long breath-holding time of ~17s (2). Recently, a deep learning (DL)-based approach (MyoMapNet) sought to shorten MOLLI imaging time to ~4s by only using four T₁-weighted images but not to improve T₁ accuracy due to using MOLLI T₁ as reference (3). This study sought to improve MyoMapNet's T₁ accuracy by using training data with ground truth T₁.

Methods

Figure 1 summarizes our study. An accelerated T₁ mapping sequence collects only four T₁-weighted images after an inversion pulse (i.e., LL4). Numerical simulations and phantom data of LL4 were used to train MyoMapNet, which is a fully connected neural network, to predict T₁. Trained MyoMapNet was evaluated by phantom and in-vivo data of LL4 and compared to SASHA and MOLLI.All MR imaging was performed on a 3T scanner (MAGNETOM Vida; Siemens Healthineers, Erlangen, Germany). MyoMapNet was implemented using PyTorch (Facebook, Menlo Park, California), trained and tested on a DGX-1 workstation (NVIDIA Santa Clara, California, United States). All simulations and T₁ calculations were performed in MATLAB 2018b (MathWorks, Natick, MA, USA).
Training Dataset
Phantom signals were collected from eighteen gel phantoms using LL4 and inversion-recovery spin-echo (IR-SE) sequence. IR-SE was used to obtain reference values. LL4 and IR-SE were performed at three different locations with the same field-of-view. LL4 was scanned 121 times, with changing simulated heart rate, flip angle, and center frequency.
Bloch equation simulations of LL4 were used to extend the training dataset to include the possible T₁, T₂, heart rate, flip angle, off-resonance frequency, noise. LL4 was simulated 5,000,000 times. Each simulation was performed with random T₁, T₂, heart rate, off-resonance, flip angle, and SNR.
Training
We used 90% of the datasets for training and 10% for validation. MyoMapNet was trained with 3000 epochs using an ADAM optimizer and a learning rate of 0.001. The batch size was 20 and rectified linear activation (Relu) was used. Parameters of MyoMapNet were updated using mean absolute error ($$$MAE=\frac{1}{number\;of\; pixels}\left|MyoMapNet\; T_1 - Ref\;T_1\right|$$$). MAE of the training datasets and validation datasets were used to select the model with the best performance.
Evaluation
Eleven NiCl2 doped agarose gel phantoms, separate from the one used in training, were scanned by SASHA, MOLLI5(3)3, and LL4 under different heart rates, center frequencies, and flip angles. Reference T₁ and T₂ values of each vial were determined IR-SE and CPMG-SE sequences, respectively. The human subject study included 19 patients (12 male; 58±16 yrs) and six healthy subjects (2 male; 23±2 yrs), which was approved by the institutional review board. Written informed consent was obtained before imaging. Native T₁ mapping was imaged in all 25 subjects, and post-contrast T₁ mapping was performed in only 16 subjects after injection of 0.1 mmol/kg Gd-DTPA (Gadavist, Bayer Healthcare, Berlin, Germany). Each subject was scanned by MOLLI, LL4, and SASHA. T₁ for individual phantom, LV and septal myocardium, and blood was measured with manually delineated ROI. ECV was calculated with blood hematocrit. Paired Student’s t-test was used in the comparison. A P-value less than 0.05 was considered statistically significant.

Results

Phantom Experiment
Phantom T₁ times measured by all three sequences were strongly correlated with reference values (Figures 2A-C) and were less sensitive to the heart rate (Figures 2D-F). As expected, T₁ underestimation of MOLLI5(3)3 was increased along with increasing of off-resonance frequency and flip angle. In contrast, SASHA and MyoMapNet were insensitive to these two confounders.
In-Vivo Study
Representative maps showed similar image quality for MyoMapNet and MOLLI, with increased noise in SASHA images (Figures 3 and 4). In Figure 5, native T₁ of the myocardium by MyoMapNet was ~50 ms lower than that by SASHA, but ~250 ms higher than that by MOLLI5(3)3 (all P<0.05). Native blood T₁ from MyoMapNet was slightly higher than that by SASHA and MOLLI5(3)3 (~60 ms and 80 ms). Post-contrast T₁ of the myocardium by MyoMapNet was ~90 ms lower than that by SASHA but ~80 ms higher than that by MOLLI4(1)3(1)2 (all P<0.05). Post-contrast blood T₁ from MyoMapNet was ~30 ms and 50 ms higher than those from SASHA and MOLLI (all P<0.05), respectively. MyoMapNet had a slightly higher SD in both native and post-contrast T₁ than MOLLI but smaller than SASHA. Mean ECV was ~22% by SASHA, 30% by MOLLI, and 31% by MyoMapNet (All P<0.05 except septum ECV between MOLLI and MyoMapNet), respectively.

Discussion and Conclusion

The improved MyoMapNet, trained using accurate T₁ values, allowed accelerating myocardial T₁ mapping completed within four heartbeats had high T₁ accuracy compared to MOLLI and was less sensitive to the off-resonance, flip angle, and heart rate as SASHA. Improved T₁ of MyoMapNet had slightly low T₁ precision compared to MOLLI but was better than SASHA. MyoMapNet T₁ is still impacted by the in-flow effect, fat, T₂, magnetic transfer, and motion artifacts as there were no in-vivo signals with ground-truth T₁ times for training. Further investigation is warranted to explore the alternative DL models for MyoMapNet to improve precision.

Acknowledgements

No acknowledgement found.