INTRODUCTION

Cardiovascular magnetic resonance T₁ mapping could characterize either diffuse or focus changes in the myocardium. Therefore, accurate myocardial T₁ measurement is an unmet need to differentiate heart diseases and facilitate multi-center and long-term studies[1]. Saturation-recovery-single-shot-acquisition (SASHA) with a three-parameter (3-Para) fitting model is an alternative cardiac T₁ mapping to Modified-Look-Locker inversion-recovery (MOLLI), which is widely used but has low accuracy, to accurately measure myocardial T₁[2,3]. Due to the application of the saturation-recovery scheme, SASHA T₁-weighted images are more prone to noise, resulting in low precision and therefore hampering the clinical application[4]. Several solutions have been proposed to improve SASHA T₁ precision, including using a variable-flip-angle sequence or a 2-parameter (2-Para) fitting model[4,5]. The variable-flip-angle sequence is a bit complex and 2-Para fitting losses accuracy. Hence, SASHA precision still needs to be improved. In this study, we developed a 1D neural network (DeepFittingNet) to predict SASHA T₁ and alleviate the impaction from noise.

METHODS

DeepFittingNet
In Figure 1, DeepFittingNet consists of a bidirectional recurrent neural network (RNN) to encode the signals (S_i) and saturation-recovery time (t_i) and a fully connected neural network (FCNN) to predict T₁. RNN and FCNN have eight hidden layers. A leaky rectified linear unit activation function is applied to each hidden layer of RNN and FCNN.
Training Dataset
SASHA sequence was numerically simulated 5000000 times. T₁ (from 100ms to 2500ms), T₂ (from 20ms to 250ms), heart rate (from 30bpm to 130bpm), off-resonance frequency (0Hz±30Hz), flip angle (Normalized B₁:1±0.2), inversion efficiency (80%±20%), and saturation efficiency (100%±20%) were randomly changed. Reference T₁ for each signal time course was calculated using a 3-Para curve-fitting method. Then, independent Gaussian noise was respectively added to real and imaginary parts of the complex singles to simulate different SNRs (from 20 to 120). Signals with and without noise were included in the dataset as data augmentation.
Validation and Testing Dataset: Twenty-one subjects (15 male;28±9 yrs) were imaged by MOLLI3(3)3(3)5 and SASHA on a Philips 3T scanner (Achieva TX) using typical imaging parameters, which was approved by the Institutional Review Board. All participants provided written informed consent prior scan. Ten of them (7 male; 29±12 yrs) were scanned twice in two separate scans. For each subject, only the SAX mid-ventricular slice was scanned. The curve-fitting method was used to fit T₁. For SASHA, 2-Para and 3-Para models were used.
Training, Validation, and Testing: We implemented DeepFittingNet on a standard PC with one NVIDIA RTX3070 graphic processing unit using the Pytorch library.
The input T₁-weighted signals were normalized to 0-1.1. DeepFittingNet was trained with 500 epochs and 25600 batch size. Mean absolute error ($$$MAE=\frac{1}{M}\sum_{}^{}|T_{x}-\hat{T_{x}}\mid$$$ , where $$$\hat{T_{x}}$$$ and $$$T_{x}$$$ are predicted one and ground-truth value) was used as the loss function to update DeepFittingNet with an ADAM optimizer and a learning rate of 0.001. SASHA T₁ maps from ten subjects were used in the validation. Rest in-vivo images were used in testing. During training, the mean absolute difference and SD of left ventricle T₁ of validation data were calculated.
Image and Statistical Analysis
ROIs for the left ventricle, septum, and blood were manually segmented in curve-fitting T₁ maps. Mean±SD of T₁ within these regions were calculated for each map. Paired Student’s t-test method with Bonferroni correction was used for multiple comparisons of the SASHA 3 T₁ calculation methods. P<0.0167 was considered statistically significant.

RESULTS

In Figure 2, SASHA T₁ maps by 2-Para fitting and DeepFittingNet had less variation across the left-ventricle myocardium in comparison to 3-Para fitting (Figure 2). MOLLI3(3)3(3)5 T₁ maps had the highest image quality.
T₁ of the left ventricle of healthy volunteers at 3T were ~1580 ms by SASHA and ~1120 ms by MOLLI3(3)3(3)5 (Figure 3 and Table 1). The corresponding blood T₁ was ~2200 ms and ~1800 ms, respectively. SASHA T₁ of the left ventricle and septum calculated by 3-Para fitting and DeepFittingNet were ~80 ms higher than those by 2-Para fitting. There is no difference in SASHA left-ventricle T₁ between 3-Para fitting and DeepFittingNet (P=0.25), as well as septum T1 (P=0.94), meaning DeepFittingNet had comparable accuracy to 3-Para fitting.
SASHA T₁ by DeepFittingNet had a lower SD than 3-Para fitting in both the left ventricle and septum, where 2-Para had the lowest value. There is no difference in SD of Blood T₁ between DeepFittingNet and 2-Para (P=0.78) while 3-Para fitting still had the highest value, indicating DeepFittingNet improved the image quality of SASHA T₁ maps in comparison with 3-Para fitting.

DISCUSSION and CONCLUSION

Our study demonstrated that a 1D neural network trained by simulated signals with different Rician noise levels could improve the SASHA T₁ precision while maintaining accuracy.