Motion artifact reduction of magnetic resonance images with an adversarial trained network

Patent metadata

Field	Value
Publication number	US-10698063-B2
Application number	US-201816008086-A
Country	US
Kind code	B2
Filing date	Jun 14, 2018
Priority date	Nov 1, 2017
Publication date	Jun 30, 2020
Grant date	Jun 30, 2020

Abstract

Systems and methods are provided for correcting motion artifacts in magnetic resonance images. An image-to-image neural network is used to generate motion corrected magnetic resonance data given motion corrupted magnetic resonance data. The image-to-image neural network is coupled within an adversarial network to help refine the generated magnetic resonance data. The adversarial network includes a generator network (the image-to-image neural network) and a discriminator network. The generator network is trained to minimize a loss function based on a Wasserstein distance when generating MR data. The discriminator network is trained to differentiate the motion corrected MR data from motion artifact free MR data.

First claim

The invention claimed is:

1. A method for reducing motion artifact in a magnetic resonance imaging system, the method comprising:
scanning a patient by the magnetic resonance imaging system to acquire magnetic resonance data;
inputting the magnetic resonance data to a machine learnt generator network configured as a fully convolutional DenseNet, the machine learnt generator network trained as an image-to-image network to generate motion corrected magnetic resonance data given motion corrupted input data;
generating, by the machine learnt generator network, first motion corrected magnetic resonance data from the input magnetic resonance data; and
displaying the first motion corrected magnetic resonance data as an image.

2. The method of claim 1 , wherein the machine learnt generator network is trained by:
generating, by the machine learnt generator network, second motion corrected magnetic resonance data from motion corrupted magnetic resonance data;
determining, using a discriminator network, how likely the second motion corrected magnetic resonance data is motion artifact free magnetic resonance data acquired by the magnetic resonance imaging system or how likely the second motion corrected magnetic resonance data was generated by the machine learnt generator network; and
adjusting the machine learnt generator network as a function of the determination.

3. The method of claim 2 , wherein how likely is quantified by:
generating a value function as a comparison between a first probability distribution of values for the second motion corrected magnetic resonance data with a second probability distribution of values for motion artifact free magnetic resonance data acquired by the magnetic resonance imaging system.

4. The method of claim 3 , wherein the value function is calculated as a function of a Wasserstein distance between the first probability distribution of values and the second probability distribution of values.

5. The method of claim 2 , wherein the discriminator network is a patch discriminator network.

6. The method of claim 2 , wherein the training further comprises:
adjusting the discriminator network to make the determination more accurate.

7. The method of claim 2 , wherein adjusting comprises:
altering weights of one or more filters of the generator network.

8. The method of claim 1 , wherein the first motion corrupted magnetic resonance data includes one or more motion artifacts.

9. The method of claim 8 , wherein the one or more motion artifacts include at least motion blurring.

10. The method of claim 1 , further comprising:
evaluating the magnetic resonance data to determine a level of motion artifacts; wherein magnetic resonance data with low levels of motion artifacts is not input into the machine learnt generator network.

11. A method for training a first neural network to generate motion corrected magnetic resonance data, the method comprising:
inputting first motion corrupted magnetic resonance data to the first neural network;
generating first motion corrected magnetic resonance data by the first neural network;
inputting the first motion corrected magnetic resonance data to a second neural network;
evaluating, by the second neural network, the first motion corrected magnetic resonance data as a function of a Wasserstein distance between a first probability distribution of the first motion corrected magnetic resonance data and a second probability distribution of a plurality of ground truth magnetic resonance data;
adjusting the first neural network based on the evaluation;
inputting second motion corrupted magnetic resonance data to the first neural network; and
repeating generating, inputting, scoring, and adjusting until the second neural network is unable to distinguish an input motion corrected magnetic resonance data from an image of the plurality of ground truth magnetic resonance data.

12. The method of claim 11 , wherein the first neural network is a convolutional DenseNet and the second neural network is a patch network.

13. The method of claim 11 , wherein the first motion corrupted magnetic resonance data includes one or more motion artifacts.

14. The method of claim 13 , wherein the one or more motion artifacts include at least ghosting.

15. A system for motion artifact correction of magnetic resonance images, the system comprising:
a magnetic resonance system configured to acquire magnetic resonance data for a patient, wherein the magnetic resonance data includes at least one or more motion artifacts;
a generator network configured as a fully convolutional DenseNet trained with a discriminator network configured as a patch discriminator and a critic function based on a Wasserstein distance to generate motion corrected magnetic resonance data when the magnetic resonance data includes one or more motion artifacts, wherein the discriminator network provides feedback to the generator network during training as a function of a Wasserstein distance between probability distributions;
a memory configured to store the generator network as trained; and
a display configured to display the motion corrected magnetic resonance data from the generator network.

16. The system of claim 15 , wherein the one or more motion artifacts include at least motion blurring.

17. The system of claim 15 , further comprising:
a filter configured to evaluate the magnetic resonance data for a level of motion artifacts.

Assignees

• Siemens Healthcare Gmbh

Inventors

• Braun Sandro
• Mailhe Boris
• Chen Xiao
• Odry Benjamin L
• Ceccaldi Pascal
• Nadar Mariappan S

Classifications

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/094

  Adversarial learning · CPC title