DEEP LEARNING TECHNIQUES FOR SUPPRESSING ARTEFACTS IN MAGNETIC RESONANCE IMAGES

1 . A method, comprising:

obtaining input magnetic resonance (MR) data using at least one radio-frequency (RF) coil of a magnetic resonance imaging (MRI) system; and

generating an MR image from the input MR data at least in part by suppressing artefacts in the input MR data, wherein the generating comprises:

performing a reconstruction step by generating an image in an image domain from the input MR data;

using a first neural network portion to suppress a first type of artefact in the input MR data in a domain other than the image domain before the reconstruction step; and

using a second neural network portion to suppress a second type of artefact in the input MR data in the domain other than the image domain before the reconstruction step.

2 . The method of claim 1 , wherein the first type of artefact comprises RF interference, and the second type of artefact comprises noise.

3 . The method of claim 1 , wherein the generating further comprises using the neural network model to suppress noise in the input MR data by processing the input MR data before the reconstruction step in the domain other than the image domain, wherein the noise is generated by a circuitry in an MR receiver chain of the MRI system and/or by a subject or object being imaged by the MRI system.

4 . The method of claim 2 , wherein the RF interference comprises external RF interference generated by a device external to the MRI system.

5 . The method of claim 1 , wherein the input MR data is in a spatial frequency domain or a sensor domain.

6 . The method of claim 1 ,

wherein the neural network model comprises a spectral unpooling layer, and

wherein using the first neural network portion to suppress the first type of artefact in the input MR data comprises applying the spectral unpooling layer to the input MR data.

7 . The method of claim 6 , wherein the neural network model further comprises a spectral pooling layer, a plurality of convolutional layers, and a skip connection.

8 . The method of claim 6 , wherein applying the spectral unpooling layer comprises applying a pointwise multiplication layer for combining first features having a first resolution provided via a skip connection with second features having a second resolution lower than the first resolution.

9 . The method of claim 1 , further comprising:

obtaining, during a first time period, RF artefact measurements using the at least one RF coil of the MRI system, wherein the RF artefact measurements include measurements of at least one of RF interference or noise; and

obtaining, during a second time period different from the first time period, MR measurements of a subject in the imaging region of the MRI system.

10 . The method of claim 9 , further comprising:

generating artefact-corrupted MR data by combining the RF artefact measurements with the MR measurements of the subject; and

training the neural network model using the artefact-corrupted MR data.

11 . The method of claim 1 , further comprising:

synthesizing RF artefact measurements, wherein the RF artefact measurements include synthesized measurements of at least one of: RF interference or noise; and

obtaining MR measurements of a subject in the imaging region of the MRI system.

12 . The method of claim 11 , further comprising: generating artefact-corrupted MR data by combining the synthesized RF artefact measurements with the MR measurements of the subject; and

training the neural network model using the artefact-corrupted MR data.

13 . A magnetic resonance imaging (MRI) system, comprising:

at least one radio-frequency (RF) coil; and

at least one processor configured to:

obtain input magnetic resonance (MR) data using the at least one RF coil; and

generate an MR image from the input MR data at least in part by using a neural network model to suppress artefacts in the input MR data, comprising:

performing a reconstruction step by generating an image in an image domain from the input MR data;

using a first neural network portion to suppress the a first type of artefact in the input MR data in a domain other than the image domain before the reconstruction step; and

using a second neural network portion to suppress a second type of artefact in the input MR data in the domain other than the image domain before the reconstruction step.

14 . The MRI system of claim 13 , wherein the first type of artefact comprises RF interference, and the second type of artefact comprises noise.

15 . The MRI system of claim 13 , wherein the generating further comprises using the neural network model to suppress noise in the input MR data by processing the input MR data before the reconstruction step in the domain other than the image domain, wherein the noise is generated by a circuitry in an MR receiver chain of the MRI system and/or by a subject or object being imaged by the MRI system.

16 . MRI system of claim 14 , wherein the RF interference comprises external RF interference generated by a device external to the MRI system.

17 . MRI system of claim 13 , wherein the input MR data is in a spatial frequency domain or a sensor domain.

18 . MRI system of claim 13 ,

wherein the neural network model comprises a spectral unpooling layer, and

wherein using the first neural network portion to suppress the first type of artefact in the input MR data comprises applying the spectral unpooling layer to the input MR data.

19 . MRI system of claim 18 , wherein the neural network model further comprises a spectral pooling layer, a plurality of convolutional layers, and a skip connection.

20 . MRI system of claim 13 , wherein applying the spectral unpooling layer comprises applying a pointwise multiplication layer for combining first features having a first resolution provided via a skip connection with second features having a second resolution lower than the first resolution.