Multi-contrast MRI image reconstruction using machine learning

Patent metadata

Field	Value
Publication number	US-12027254-B2
Application number	US-202117179999-A
Country	US
Kind code	B2
Filing date	Feb 19, 2021
Priority date	Feb 21, 2020
Publication date	Jul 2, 2024
Grant date	Jul 2, 2024

Abstract

A method for reconstructing a MRI image may include: receiving MRI measurement data sets f 1 to f N , each data set being acquired from an examination object based on a different MRI protocol of an MRI system; receiving MRI images u 1 0 to u N 0 corresponding to the MRI measurement data sets f 1 to f N ; performing one or more translation and rotation transformations on the MRI images u 1 0 to u N 0 ; applying one or more trained functions: to the transformed MRI images u 1 0 to u N 0 , using a neural network, and to the MRI images u 1 0 to u N 0 , using a forward-sampling operator; performing one or more inverse translation and rotation transformations on an output of the neural network; and generating at least one output MRI image u T based on an output of the forward-sampling operator, the inversely transformed output of the neural network, and the input MRI images u 1 0 to u N 0 .

First claim

The invention claimed is:

1. A computer-implemented method for reconstructing a magnetic resonance imaging (MRI) image, comprising:
receiving a plurality of MRI measurement data sets f 1 to f N , wherein each data set is acquired from an examination object based on a different MRI protocol of an MRI system;
receiving MRI images u 1 0 to u N 0 corresponding to the MRI measurement data sets f 1 to f N ; and
in at least a first step GD 1 :
performing one or more translation and rotation transformations on the MRI images u 1 0 to u N 0 ;
applying one or more trained functions (i) to the transformed MRI images u 1 0 to u N 0 , using a neural network, and (ii) to the MRI images u 1 0 to u N 0 , using a forward-sampling operator;
wherein an output of the neural network comprises a result of filtering the transformed MRI images u 1 0 to u N 0 via the one or more trained functions based upon a data-fidelity term that is computed individually by the forward-sampling operator for each one of the MRI images u 1 0 to u N 0 ;
performing one or more inverse translation and rotation transformations on the output of the neural network to transform each of the MRI images u 1 0 to u N 0 back to an original image orientation to preserve agreement with the forward-sampling operator;
generating at least one output MRI image u T based on an output of the forward-sampling operator, the inversely transformed output of the neural network, and the input MRI images u 1 0 to u N 0 ; and
providing the at least one output MRI image u T .

2. The computer-implemented method according to claim 1 , wherein the forward-sampling operator determines an agreement between at least one MRI image u 1 0 to u N 0 and the corresponding MRI measurement data set f 1 to f N .

3. The computer-implemented method according to claim 1 , wherein applying trained functions to the MRI images u 1 0 to u N 0 comprises:
applying an unrolled model to the MRI images u 1 0 to u N 0 , wherein, for each input MRI image u 1 0 to u N 0 , an output of the neural network and an output of the forward-sampling operator is subtracted from the respective input MRI image u 1 0 to u N 0 to generate the at least one output MRI image u T .

4. The computer-implemented method according to claim 1 , wherein the neural network is a convolutional neural network, and wherein applying trained functions to the MRI images u 1 0 to u N 0 comprises:
applying a plurality of trained filter functions to the MRI images u 1 0 to u N 0 simultaneously, using convolutions and non-linear activations, wherein each filter function is applied to each MRI image u 1 0 to u N 0 .

5. The computer-implemented method according to claim 1 , wherein, for each of the MRI images u 1 0 to u N 0 separately, the forward-sampling operator determines an agreement between each of the MRI images u 1 0 to u N 0 and the corresponding MRI measurement data sets f 1 to f N using a MRI forward model of the corresponding MRI protocol.

6. The computer-implemented method according to claim 1 , further comprising:
receiving coil sensitivities C of the MRI system; and
receiving an under-sampling scheme for each MRI protocol,
wherein the forward-sampling operator determines an agreement between a respective MRI image u 1 0 to u N 0 and the corresponding MRI measurement data set f 1 to f N using a MRI forward model of the MRI protocol, based on the coil sensitivities C and the respective under-sampling scheme.

7. The computer-implemented method according to claim 1 , wherein applying the trained functions to the MRI images u 1 0 to u N 0 comprises:
applying, in a plurality of steps GD 1 to GD T for a predefined number T, trained functions to the MRI images u 1 0 to u N 0 , each step GD t for t=1 . . . T comprising:
receiving input MRI images u 1 t−1 to u N t−1 ;
applying trained functions: to transformations of the MRI images u 1 t−1 to u N 1−1 , using the neural network, to generate the output of the neural network, and to the MRI images u 1 t−1 to u N 1−1 , using the forward-sampling operator, to generate the output of the forward-sampling operator, wherein MRI output images u i t to u N t are generated based on the output of the neural network and the output of the forward-sampling operator; and
providing the MRI output images u i t to u N t .

8. The computer-implemented method according to claim 7 , wherein generating the MRI output images u 1 t to u N t comprises:
subtracting, for each input MRI image u 1 t−1 to u N 1−1 , an corresponding output of the neural network and output of the forward-sampling operator from the input MRI image u 1 t to u N t .

9. The computer-implemented method according to claim 1 , wherein the forward-sampling operator determines an agreement between each MRI image u i t−1 , with i=0 . . . N, and the corresponding MRI measurement dataset f i , using the relation:
λ i t−1 A* i (A i u i t−1 −f i ),
wherein λ i t−1 is a contrast specific regularization parameter, A i is a MRI forward model for the corresponding MRI protocol, and A* i is an adjoint of the MRI forward model A i .

10. The computer-implemented method according to claim 1 , wherein the MRI measurements data sets f 1 to f N are based on different MRI contrasts of the MRI system.

11. The computer-implemented method according to claim 1 , wherein the MRI measurement data sets f 1 to f N are acquired based on MRI protocols with different contrasts.

12. The computer-implemented method according to claim 1 , wherein the at least one output MRI image u T is an output image u i T , for i={1 . . . N}, of a plurality of output MRI images u 1 T to u N T , wherein each output MRI image u 1 T to u N T corresponds to a respective one of the MRI images u 1 0 to u N 0 .

13. The computer-implemented method according to claim 1 , wherein each MRI measurement data set f 1 to f N is undersampled in k-space based on a specific under-sampling scheme.

14. The computer-implemented method according to claim 9 , wherein at least two of the MRI measurement data sets f 1 to f N are based on different under-sampling schemes.

15. The computer-implemented method according to claim 1 , wherein reconstructing MRI images u 1 0 to u N 0 from the MRI measurement data sets f 1 to f N further comprises:
removing a low-resolution background phase from the MRI images u 1 0 to u N 0 .

16. The computer-implemented method according to claim 1 , wherein applying trained functions to the MRI images u 1 0 to u N 0 further comprises:
dividing coil sensitivities C of the MRI system, the MRI measurement data sets f 1 to f N , the MRI images u 1 0 to u N 0 , and reference images g 1 to g N into patches of collapsing voxel groups, and
applying the trained functions to the patches of collapsing voxel groups of each MRI image u 1 0 to u N 0 individually.

17. The computer-implemented method according to claim 1 , wherein reconstructing MRI images u 1 0 to u N 0 further comprises registering initial parallel imagining reconstructions of the MRI images u 1 0 to u N 0 to determine motion parameters associated with inter-scan motion to increase co-registration between jointly reconstructed scans.

18. A computer configured for reconstructing an MRI image, the computer comprising:
a memory storing executable instructions;
an interface; and
a processor configured to exec

Assignees

• Siemens Healthcare Gmbh
• Massachusetts Gen Hospital
• Siemens Healthineers Ag

Inventors

• Polak Daniel
• Setsompop Kawin

Classifications

• A61B5/055

  involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging · CPC title

• G06T12/20

  Inverse problem, i.e. transformations from projection space into object space · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06V10/82

  using neural networks · CPC title