Purpose:

The aim of this work is to provide a dictionary-less approach to reconstruct multi-parametric MR maps based on Tailored Magnetic Resonance Fingerprinting acquisition method.

Introduction:

Magnetic Resonance Fingerprinting (MRF) is an accelerated acquisition and reconstruction technique employed to generate multiple parametric maps¹ Tailored MRF (TMRF) has been proposed with a block based (T₁, T₂ and PD) acquisition² approach to overcome the limitation of under estimation of long T₂ by dictionary resolution. To overcome this limitation, a deep learning based approach is proposed for the reconstruction of TMRF data. A three-layer Deep Neural Network (DNN) created on TensorFlow³ is used for training.

Methods:

The reconstruction pipeline for tailored MRF using block based acquisition is demonstrated in Figure 1. Acquisition: The signal intensity of a gradient echo based sequence is more dependent on the Flip Angle (FA) than the Repetition Time (TR) due to the minimal TRs typically employed in such cases⁴. Thus, the required contrast can be achieved by an optimal choice of the FA. This has been utilized to form 3 blocks that optimize contrast for Proton Density (PD), T₁ and T₂ weighting. This allows each tissue type to have hyper-intensities in at least one of the three blocks. TRs and FAs were independently designed for each block and then combined into a single sequence. Each block comprised of 240 acquisitions (total of 720 TR/FA combinations). Four human in vivo brain data were acquired on a 1.5 T GE Signa with a spiral readout time of 5ms with a fixed Echo Time (TE) of 2.7ms, as part of an institution approved study. The spiral trajectory consisted of 48 arms and 720 images acquired using the TR/FA schedule was sliding window reconstructed to get 673 images. Training: We considered two approaches for general/brain specific image reconstruction. Approach 1: 10000 Natural images of size 32x32 were downloaded from CIFAR-1005⁵ for training the NN. To improve the robustness of the proposed approach, 50% of the data was corrupted by introducing three artifacts: random rotation, Rician noise and circular shift because these discrepancies were commonly seen in TMRF data. As input to the Neural Network, 10000 voxels were randomly selected from synthesized data of size 32x32x673x10000 and trained against concurrent voxels selected from GT maps (T₁ and T₂). Approach 2: Synthetic brain images of size 128x128 were synthesized using the above method to create a training dataset of size 128x128x673. Similar to the previous approach, 12000 voxels were selected from the synthesized maps and GT maps (synthetic brain images) for NN training. The voxels selected were non-zero values selected from the region of the brain. DNN: T₁ and T₂ for approaches 1 and 2 were separately trained. A three layered DNN was used for both parameters. The DNN for T₂ had weights of size 64, 32 and 1 for the layers, whereas the DNN for T₁ had weights of size 128, 64 and 1. Testing: Voxel based testing was performed for both approaches on the data acquired using TMRF method. Four such datasets were used to test the DNN. The data was forward propagated through the trained NN to get the reconstructed voxels. These voxels were reshaped to generate reconstructed maps on Matlab (The Mathworks Inc, MA). The reconstructed maps were compared with the scanner generated GT maps to validate the proposed approaches. The code for this implementation is available online⁶.

Discussion and Conclusion

It can be ascertained that the results obtained from the first better compared to the second. We attribute this to the overfitting nature of the second neural network. This is also shown by the L-curves for training. A salient feature of the natural images approach is that it is not restricted to a single organ.

Acknowledgements

This work was supported by Department of Science and Technology (DST) – DST/TSG/NTS/2013/100 and Vision Group on Science and Technology, Govt. of Karnataka, Karnataka Fund for strengthening infrastructure (K-FSIT), GRD#333/2015.