Artificial intelligence based suppression of chemical species in magnetic resonance imaging
A computer-implemented method for using machine learning to suppress fat in acquired MR images includes receiving multi-echo images from an anatomical area of interest acquired using an MRI system. A first subset of the multi-echo images is acquired prior to application of contrast to the anatomical area of interest and a second subset of the multi-echo images is acquired after application of contrast to the anatomical area of interest. Next, data is generated including water images, fat images, and effective R* 2 maps from the multi-echo images. The water images, the fat images, and the effective R* 2 maps are used to create synthetic fat suppressed images. A neural network is trained to use the multi-echo images as input and the synthetic fat suppressed images as ground truth. A plurality of components of the neural network are saved to allow later deployment of the neural network on a computing system.
1. A computer-implemented method for using machine learning to suppress fat in acquired magnetic resonance (MR) images, the method comprising:
receiving a plurality of multi-echo images from an anatomical area of interest acquired using an MRI system, wherein a first subset of the multi-echo images is acquired prior to application of contrast to the anatomical area of interest and a second subset of the multi-echo images is acquired after application of contrast to the anatomical area of interest;
generating a plurality of water images, a plurality of fat images, and a plurality of effective R* 2 maps from the plurality of multi-echo images;
using the water images, the fat images, and the effective R* 2 maps to create a plurality of synthetic fat suppressed images;
training a neural network to use the multi-echo images as input and the synthetic fat suppressed images as ground truth; and
saving a plurality of components of the neural network to allow later deployment of the neural network on a computing system.
2. The method of claim 1 , wherein the plurality of water images and the plurality of fat images correspond to an echo time (TE) equal to 0.
3. The method of claim 2 , wherein each synthetic fat suppressed image corresponds to a distinct TE value greater than 0.
4. The method of claim 1 , wherein the synthetic fat suppressed images are generated using multi-echo Dixon reconstruction and multi-step adaptive fitting.
5. The method of claim 1 , wherein the neural network is a deep learning network.
6. The method of claim 5 , wherein the deep learning network is based on a U-Net structure.
7. The method of claim 5 , wherein the plurality of multi-echo images are acquired in a plurality of dimensions and the deep learning network transforms the plurality of dimensions into a single virtual dimension.
8. The method of claim 7 , wherein the plurality of dimensions correspond to echoes, slices, subjects, pre-contrast processing, and post-contrast processing.
9. The method of claim 8 , wherein the plurality of dimensions further include channels of the MRI system used in acquiring the multi-echo images.
10. The method of claim 1 , wherein the plurality of components of the neural network comprise a structural description of the neural network, one or more hyper-parameters used in training the neural network, and one or more weights resulting from training of the neural network.