The bipolar multi-echo gradient recalled echo (bmGRE) imaging collects multi-echo signals in a single TR. The bmGRE has several bipolar issues from the gradient switching and the opposite gradient polarities which prevent accurate water-fat separation. The proposed methods eliminate the bipolar issues in two ways.
First, the water-fat separation is performed separately for each gradient polarity, and the results of the positive and negative gradient polarities are combined after the water-fat separation. To reduce the imaging time, the data is subsampled at every echo time and every gradient polarity, and the images are reconstruct-ed from the subsampled data using the low-rank property of the bipolar acquisition.
Second, water-fat separated images were obtained using the neural network. A convolutional neural network (CNN) was designed and trained using the relationship between multi-echo images from the bmGRE and artifact-free water-fat separated images. The artifact-free water-fat separated images for training the CNN were obtained from multiple echo-time signals using iterative decomposition of water and fat with echo asymmetry and the least-squares estimation method, where multiple signals at different echo times were acquired using a single-echo gradient recalled echo sequence. Also, we propose a data augmentation method using a synthetic field for multi-echo signals including the bipolar issues, to prevent overfitting of the network and to increase the separation accuracy of the CNN. In addition, the network became more ro-bust to various cases of field inhomogeneity thanks to the data augmentation.
The phantom and in-vivo experiments were conducted using the unipolar multi-echo GRE, the bmGRE and the Soliman’s and the proposed methods. Experimental results demonstrate that the proposed method is able to obtain accurate water and fat images without bipolar issues, in shorter imaging time. Also, we trained the CNN using in-vivo knee images and tested it using in-vivo knee, head, and ankle images. In-vivo imag-ing results showed that the proposed CNN could separate water-fat images accurately. The proposed data augmentation method could prevent overfitting even with a limited number of training data, and made the method robust to magnetic field inhomogeneities.
In the study, a new water-fat separation method using the interleaved bipolar acquisition was pro-posed to reduce the issues from the bipolar gradients with a short imaging time. Moreover, the low-rank property of the bipolar acquisition was defined and utilized to estimate the full data from the subsampled data. The proposed CNN could also separate water-fat images from the bmGRE.