Introduction

Balanced steady state free precession (bSSFP) sequence has been gaining importance in clinical practices due to its high speed, high signal-to-noise ratio (SNR), and minimal distortion. However, bSSFP often suffers from well-known banding artifacts, which are related to the field strength and repetition time (TR) (1). Various methods have been introduced to suppress banding artifacts in bSSFP, but banding artifacts cannot be suppressed completely by these methods, especially when number of multiple phase cycling (PC) bSSFP datasets is small (2-4). In this study, we proposed a learning-based method to combine the bSSFP datasets with multiple PC for better banding artifact suppression. Models of multilayer perceptron (MLP), a feedforward artificial neural network (ANN), were trained to predict banding free bSSFP images based on two and four PC datasets. Then, the trained MLP models were applied to another bSSFP datasets for evalulation in terms of performance of banding artifact suppression.

Materials and Methods

MLP consists of three layers of input, hidden, and output layers. In training stage, inputs and outputs of the MLP models were two or four PC datasets and banding-free images, respectively. Banding-free bSSFP images were generated by maximum-intensity projection (MIP) of 8 or 12 phase-cycled datasets. In test stage, the learned MLP models received two (0 and 180°) or four (0, 90, 180, and 270°) PC datasets as input and produced banding-suppressed images as output. The training and test were performed in two ways. First, Bloch equation simulations were performed to generate 12 PC datasets. The ground truth was reconstructed by MIP of all PC datasets, among which two (0 and 180°) PC datasets were used as the input. The MLP models were trained by the inputs and the ground truth, and the trained models were evaluated with simulated ‘bi-tissue’ datasets with T1/T2= 420/50 ms and 1100/70 ms. Two separate 3D bSSFP datasets with multiple PCs were acquired in humans; axial brain bSSFP images (dataset #1) in one subject (subject #1) using 8 PCs and coronal knee bSSFP images (dataset #2) in another subject (subject #2) with 12 PCs. Using all PC datasets, the ground truth images and the inputs were produced in the same way as the simulation. To test the trained MLP models, axial brain images (dataset #3), sagittal knee images (dataset #4), and axial knee images (dataset #5) were acquired. The MLP models were trained by two or four PC datasets as input and were additionally evaluated in human experiments. Image parameters of all the scans were shown in Table 1. The mean square error (MSE) index was used to measure the difference between the ground truth images and the outputs of MLP, MIP, and sum-of-square (SOS), and SNR was also calculated. All experiments were performed with a 3T whole-body scanner (Siemens Medical Solutions, Erlangen, Germany). Image analysis was performed using MATLAB(The Mathworks, Natick, MA).

Results

The simulation showed that MLP was superior to the existing methods (MIP and SOS) for banding artifacts suppression (Figure 1): The images combined with MIP and SOS demonstrated visible signal fluctuations, which were mostly eliminated in the results of MLP. MSE values of MLP (1.5X10^-5) were much smaller than those of the existing methods (MIP: 5.6x10^-5, SOS: 6.7x10^-4). The results of human experiments were consistent with those of simulations. MLP showed lower MSE values than those of MIP in both training datasets and test datasets (Table 2), except for the four PC images of dataset #3. The bSSFP images at different PC angles showed banding artifacts at different spatial locations (Figure 2a). Visually both MIP and MLP suppressed banding artifacts well (Figure 2b-e), however, the differences between the ground truth images and MIP/MLP images were clearly visualized in the subtraction images (Figure 2g-j). MLP showed smaller difference than that of MIP. MLP showed higher SNR efficiency than MIP in all tissues (Table 3).

Discussion and Conclusion

To our knowledge, it is the first time for a learning-based algorithm to be applied to reducing banding artifacts in bSSFP. Combination of multiple PC datasets with MLP successfully suppressed banding artifacts in bSSFP MRI, while maintaining high SNR. The MLP models trained with a couple of 3D multiple PC datasets from two different organs in human worked well on datasets acquired with different scan parameters from different subjects, and also those acquired from a different species at different field strength. MLP worked well even with two PC datasets, providing better banding artifact suppression and higher SNR efficiency than the widely-used MIP. Artificial neural network is a promising approach to combining multiple phase cycled bSSFP datasets for banding artifact suppression.

Acknowledgements

No acknowledgement found.