Introduction

T1 mapping of the abdomen has been demonstrated to be useful for quantitative evaluation of various pathologies.^1,2 Recently, rapid 2D radial Look Locker techniques were proposed for multi-slice high resolution T1 mapping of the abdomen.^3,4 These techniques allow for the acquisition ~10 high-resolution slices within a single breath hold period (BHP) and therefore require multiple BHPs for full abdominal coverage. Deep-learning (DL) techniques have been proposed for MR relaxometry applications.^5,6 These DL techniques exploit the spatio-temporal redundancies to improve T1 estimation performance. In this study, we propose a DL approach to improve the slice coverage of 2D radial Look Locker methods.

Methods

Figure 1 summarizes the radial 2D Look Locker T1-mapping technique (inversion recovery radial steady-state free precession (IR-radSSFP)).¹ Let $$$T_{S}$$$ and $$$S$$$ represent the acquisition time for one slice and the number of acquired slices, respectively. The radial views acquired throughout the T1 recovery curve (T1RC) for each slice are grouped into N inversion time (TI) groups from which N TI images are reconstructed. Data acquisition for all acquired slices need to be completed within the BHP time $$$T_{BHP}$$$: $$$S\cdot T_{S}\leq T_{BHP}$$$. Therefore, to accommodate more slices, the acquisition time per slice $$$T_{S}$$$ needs to be reduced. If the $$$T_{S}$$$ is reduced substantially (e.g. less than 2 seconds) by reducing N, the T1 relaxation time cannot be estimated accurately using conventional nonlinear least squares (NLLS) curve fitting,⁷ especially for long T1 species. We note, however, that the T1 signal evolves rapidly during early recovery and slows in the later recovery regime (Figure 1.a). Thus, we propose a strategy where the number of TIs along the T1RC is reduced but T1 estimation is performed using a convolutional neural network (CNN), which exploits the spatio-temporal redundancies in T1RC to improve T1 estimation performance.

The proposed DL framework is presented in Figure 2. The inputs to our DL model ($$$X=\{{I_{1},...,I_{\hat{N}}}\}$$$) are a set of TI images which have undergone iterative reconstruction, where $$$\hat{N}\leq N$$$ indicates a reduced number of TIs. The output of our model $$$\hat{T}_{1}$$$ is an estimate of the T1 map. The network targets/labels are T1 maps obtained using N TI images.

Ten subjects were imaged using the 2D IR-radSSFP sequence with TR=4.40ms, TE=2.15ms, 32 TIs, in-plane resolution=0.83mmX0.83mm, slice thickness=3mm, 10 slices, and 16 lines/TI with 384 readout points/line. Reference T1 maps were obtained using a model-based CS approach⁸ followed by reduced dimension NLLS curve fitting.⁷ These full T1RC datasets were retrospectively undersampled by selecting the first $$$\hat{N}=8,12,16,20,24,28$$$ TI groups. Similarly, the TI images at different number of truncations were individually reconstructed using the CS approach. For each $$$\hat{N}$$$, a neural network is trained using $$$\hat{N}$$$ TI images as inputs and the corresponding reference T1 maps obtained with $$$N=32$$$ as targets. We used ResNet⁹ as the backbone of our network consisting of 16 residual blocks, where each block contains two 32-filter convolutional layers with batch normalization. The network was trained for 100 epochs using a fixed learning rate of 1e-4 and an L2 weight decay of 1e-4. Patch extraction (16x16) as well as random rotation and flip were applied for data augmentation. In each experiment, 7 subjects were randomly selected for training, 2 subjects were randomly selected for testing. One subject with liver pathology was reserved for testing. ROI analysis of estimated T1 values was performed for NLLS curve fitting and DL methods for each $$$\hat{N}$$$ and compared to reference T1 values.

Results and Discussion

Figure 3 demonstrates performance of the NLLS method when a varying number of TIs are used in estimation. We observe that T1 estimation performance deteriorates significantly in regions where signal-to-noise ratio (SNR) is low, due to poor curve fitting, resulting in substantial variation in T1 estimates between neighboring pixels . Figure 4 provides a comparison of T1 estimations using NLLS and the proposed DL framework. Nine ROIs (with varying noise levels and tissue types) from three slices were used to evaluate the T1 mapping performances. The DL model achieved under 2% estimation error (as measured by relative mean absolute error) across all number of TIs ($$$\hat{N}$$$) in lower-noise regions of the liver, whereas the estimation error for NLLS increased sharply below 16 TIs. For the higher-noise liver and spleen regions the DL model estimation error is within 6% across all $$$\hat{N}$$$. In contrast, NLLS yields large estimation errors when $$$\hat{N}$$$ falls below 24. Figure 5 provides an example of T1 values in a subject with a liver lesion using NLLS and DL methods for varying $$$\hat{N}$$$. Within the lesion, the DL model achieved under 6% estimation error up until $$$\hat{N}=16$$$, while for NLLS estimation error exceeded 6% after $$$\hat{N}=28$$$. Figure 5.b provides visual examples of the T1 maps of the lesion ROI, which demonstrate the superior structural integrity of the DL estimated T1 maps across $$$\hat{N}$$$ values.

Conclusions

We presented an accelerated T1 mapping framework which utilizes deep learning to extend the slice coverage of 2D radial Look Locker T1 mapping methods. In vivo experiments demonstrate that the proposed method achieves less than 6% T1-estimation error while extending the slice coverage by 4X compared to recent radial Look Locker T1 mapping techniques.

Acknowledgements

We would like to acknowledge grant support from NIH (CA245920 and CA245920S1), the Arizona Biomedical Research Commission (ADHS14-082996), and the Technology and Research Initiative Fund Technology and Research Initiative Fund (TRIF).