Reconstruction of magnetic-resonance datasets using machine learning

1. A method, comprising:

determining an initial magnetic resonance imaging (MRI) dataset in an image domain based on an initial reconstruction of MRI measurement data obtained using an undersampling scheme;

determining patches of the initial MRI dataset in accordance with a patching scheme, the patching scheme depending on the undersampling scheme;

for each one of the patches, applying a machine-learned algorithm to obtain a respective patch of a reconstructed MRI dataset, the machine-learned algorithm depending on the undersampling scheme; and

combining the patches of the reconstructed MRI dataset.

2. A system comprising:

a memory that stores program code; and

a processor configured to load the program code from the memory and execute the program code to:

determine an initial magnetic resonance imaging (MRI) dataset in an image domain based on an initial reconstruction of MRI measurement data obtained using an undersampling scheme;

determine patches of the initial MRI dataset in accordance with a patching scheme, the patching scheme depending on the undersampling scheme;

for each one of the patches, apply a machine-learned algorithm to obtain a respective patch of a reconstructed MRI dataset, the machine-learned algorithm depending on the undersampling scheme; and

combine the patches of the reconstructed MRI dataset.

3. The method of claim 1 , wherein each patch comprises voxels in multiple sub-regions of the initial MRI dataset, the sub-regions being selected in accordance with the undersampling scheme and being offset from each other.

4. The method of claim 3 , wherein the sub-regions of two patches of the initial MRI dataset are interleaved in accordance with an acceleration factor of the undersampling scheme.

5. The method of claim 3 , wherein the sub-regions of a given patch of the initial MRI dataset are shifted with respect to each other in accordance with a shifting factor of the undersampling scheme.

6. The method of claim 1 , wherein the patching scheme defines overlaps between adjacent ones of the patches.

7. The method of claim 1 , further comprising: populating the patches with voxels of the initial MRI dataset along a readout direction.

8. The method of claim 7 , wherein the patches are populated with voxels from slices of the initial MRI dataset selected in accordance with a point spread function associated with the undersampling scheme.

9. The method of claim 8 , wherein the patches are populated with voxels along the readout direction based on a sparsity of the point spread function.

10. The method of claim 1 , wherein the undersampling scheme is associated with a controlled aliasing in parallel imaging results in higher acceleration (CAIPIRINHA) encoding or a Wave-controlled aliasing in parallel imaging (Wave-CAIPI) encoding.

11. The method of claim 1 , wherein the machine-learned algorithm is applied using a processor coupled to a memory, wherein the method further comprises: determining a size of the patches based on a size of the memory.

12. The method of claim 1 , wherein the machine-learned algorithm comprises multiple iterations of a variational model, the variational model including a forward-sampling operator being applied on the respective patch of the reconstructed MRI dataset of a given iteration and the respective patch, the forward sampling operator being associated with coil sensitivity maps, Fourier transform, and the undersampling scheme.

13. The method of claim 1 , wherein the machine-learned algorithm comprises multiple iterations of a variational model, the variational model including a regularization operator being applied on the respective patch of the reconstructed MRI dataset of a given iteration, the regularization operator being associated with at least one of a convolutional filter and a non-linear filter.

14. The method of claim 7 , wherein:

the machine-learned algorithm comprises multiple iterations of a variational model, the variational model including: a forward-sampling operator being applied on the respective patch of the reconstructed MRI dataset of a given iteration and the respective patch, and a regularization operator being applied on the respective patch of the reconstructed MRI dataset of a given iteration,

the patches are populated with the voxels from a first set of slices of the initial MRI dataset,

the regularization operator is applied to the patches upon populating the patches,

upon applying the regularization operator, voxels from a second set of slices are retained in the patches, and

upon retaining the voxels from the second set of slices, the forward-sampling operator is applied.

15. The method of claim 1 , wherein the machine-learned algorithm comprises parameters that are determined based on training using a ground-truth MRI dataset.

16. The method of claim 15 , further comprising:

obtaining an MRI ground truth dataset;

determining patches of the MRI ground truth dataset in accordance with the patching scheme; and

for at least one patch of the MRI ground truth dataset, performing a training of the machine-learned algorithm based on a comparison between the respective at least one patch of the reconstructed MRI dataset and the respective at least one patch of the MRI ground truth dataset.

17. A non-transitory computer-readable storage medium with an executable program stored thereon, that when executed, instructs a processor to perform the method of claim 1 .

18. A computer program product having a computer program which is directly loadable into a memory of a controller of the magnetic resonance imaging system, when executed by the controller, causes the magnetic resonance system to perform the method as claimed in claim 1 .

19. The system of claim 2 , further comprising an MRI scanner that is configured to perform a magnetic resonance imaging on a patient to obtain the MRI measurement data.