MRI image reconstruction from undersampled data using adversarially trained generative neural network
A method of magnetic resonance imaging acquires undersampled MRI data and generates by an adversarially trained generative neural network MRI data having higher quality without using any fully-sampled data as a ground truth. The generative neural network is adversarially trained using a discriminative neural network that distinguishes between undersampled MRI training data and candidate undersampled MRI training data produced by applying an MRI measurement function containing an undersampling mask to generated MRI training data produced by the generative neural network from the undersampled MRI training data.
1. A method of magnetic resonance imaging (MRI) comprising:
acquiring by an MRI scanner undersampled MRI data;
generating by a generative neural network generated MRI data from the undersampled MRI data, wherein the generated MRI data has higher quality than the undersampled MRI data;
wherein the generative neural network is adversarially trained without using fully-sampled training data using a discriminative neural network that distinguishes between
i) undersampled MRI training data; and
ii) candidate undersampled MRI training data produced by applying an MRI measurement function to generated MRI training data produced by the generative neural network from the undersampled MRI training data, wherein the measurement function comprises an undersampling mask.
2. The method of claim 1 wherein the undersampled MRI data, the generated MRI data, the undersampled MRI training data, and the candidate undersampled MRI training data are represented in image space.
3. The method of claim 1 wherein acquiring by the MRI scanner the undersampled MRI data comprises acquiring undersampled k-space data and mapping the undersampled k-space data to image space using zero-filled reconstruction.
4. The method of claim 1 wherein the measurement function comprises coil sensitivity maps, a Fourier transform, the undersampling mask, an inverse Fourier transform, and the coil sensitivity maps again.
5. The method of claim 1 wherein the undersampled MRI data, the generated MRI data, the undersampled MRI training data, and the candidate undersampled MRI training data are represented in k-space.
6. The method of claim 1 further comprising mapping the generated MRI data from k-space to image space by applying an inverse Fourier transform and coil sensitivity maps.
7. The method of claim 1 wherein the generative neural network comprises a data consistency block followed by convolution and ReLU blocks.
8. The method of claim 1 wherein the generative neural network comprises multiple pairs of convolution and ReLU blocks followed by a hyperbolic tangent block.
9. The method of claim 1 wherein the discriminative neural network comprises multiple pairs of convolution and ReLU blocks followed by a densely connected block.
10. The method of claim 1 wherein the undersampled MRI data, the generated MRI data, the undersampled MRI training data, and the candidate undersampled MRI training data represent 2D slices, 3D volumes, 4D datasets, or 2D slices plus a time dimension.