MRI reconstruction based on generative models

Patent metadata

Field	Value
Publication number	US-12437187-B2
Application number	US-202217891702-A
Country	US
Kind code	B2
Filing date	Aug 19, 2022
Priority date	Aug 19, 2022
Publication date	Oct 7, 2025
Grant date	Oct 7, 2025

Abstract

Deep learning-based systems, methods, and instrumentalities are described herein for MRI reconstruction and/or refinement. An MRI image may be reconstructed based on under-sampled MRI information and a generative model may be trained to refine the reconstructed image, for example, by increasing the sharpness of the MRI image without introducing artifacts into the image. The generative model may be implemented using various types of artificial neural networks including a generative adversarial network. The model may be trained based on an adversarial loss and a pixel-wise image loss, and once trained, the model may be used to improve the quality of a wide range of 2D or 3D MRI images including those of a knee, brain, or heart.

First claim

What is claimed is:

1. An apparatus, comprising:
one or more processors configured to:
obtain a reconstructed magnetic resonance imaging (MRI) image of an anatomical structure, wherein the reconstructed MRI image is generated based on under-sampled MRI data associated with the anatomical structure and using a first artificial neural network trained for generating the reconstructed MRI image based on the under-sampled MRI data; and
process the reconstructed MRI image through a second artificial neural network, wherein the second artificial neural network comprises a generator network and a discriminator network, the generator network is used to refine the reconstructed MRI image of the anatomical structure, the discriminator network is used to supervise the generator network, and a refined MRI image of the anatomical structure is generated as a result of the processing, the refined MRI image characterized by an improved sharpness over the reconstructed MRI image.

2. The apparatus of claim 1 , wherein the second artificial neural network comprises a generative adversarial network (GAN).

3. The apparatus of claim 1 , wherein the discriminator network is trained to supervise the generator network so that the refined MRI image generated by the generator network follows a probability distribution of fully-sampled MRI images of the anatomical structure.

4. The apparatus of claim 1 , wherein the second artificial neural network is trained through a training process that comprises:
obtaining a reconstructed MRI training image, wherein the reconstructed MRI training image is generated based on under-sampled MRI training data;
predicting, using an instance of the second artificial neural network, an output MRI image based on the reconstructed MRI training image;
determining an adversarial loss associated with the prediction; and
adjusting parameters of the instance of the second artificial neural network based on at least the adversarial loss.

5. The apparatus of claim 4 , wherein the training process further comprises determining an additional training loss based on the output MRI image and a ground truth MRI image, and adjusting the parameters of the instance of the second artificial neural network further based on the additional training loss.

6. The apparatus of claim 5 , wherein the parameters of the instance of the second artificial neural network are adjusted based on a weighted average of the adversarial loss and the additional training loss.

7. The apparatus of claim 6 , wherein respective weights assigned to the adversarial loss and the pixel-wise loss are adjusted during the training process to balance between a sharpness and a level of artifacts in an image produced by the second artificial neural network.

8. The apparatus of claim 5 , wherein the ground truth MRI image is a fully sampled MRI image.

9. The apparatus of claim 1 , wherein the first artificial neural network and the second artificial neural network are trained together in an end-to-end manner.

10. A method of magnetic resonance imaging (MRI) reconstruction, the method comprising:
obtaining a reconstructed magnetic resonance imaging (MRI) image of an anatomical structure, wherein the reconstructed MRI image is generated based on under-sampled MRI data associated with the anatomical structure and using a first artificial neural network trained for generating the reconstructed MRI image based on the under-sampled MRI data; and
processing the reconstructed MRI image through a second artificial neural network, wherein the second artificial neural network comprises a generator network and a discriminator network, the generator network is used to refine the reconstructed MRI image of the anatomical structure, the discriminator network is used to supervise the generator network, and a refined MRI image of the anatomical structure is generated as a result of the processing, the refined MRI image characterized by an improved sharpness over the reconstructed MRI image.

11. The method of claim 10 , wherein the second artificial neural network comprises a generative adversarial network (GAN).

12. The method of claim 10 , wherein the discriminator network is trained to supervise the generator network so that the refined MRI image follows a probability distribution of fully-sampled MRI images of the anatomical structure.

13. The method of claim 10 , wherein the second artificial neural network is trained through a training process that comprises:
obtaining a reconstructed MRI training image, wherein the reconstructed MRI training image is generated based on under-sampled MRI training data;
predicting, using an instance of the second artificial neural network, an output MRI image based on the reconstructed MRI training image;
determining an adversarial loss associated with the prediction; and
adjusting parameters of the instance of the second artificial neural network based on at least the adversarial loss.

14. The method of claim 13 , wherein the training process further comprises determining an additional training loss based on the output MRI image and a ground truth MRI image, and adjusting the parameters of the instance of the second artificial neural network further based on the additional training loss.

15. The method of claim 14 , wherein the parameters of the instance of the second artificial neural network are adjusted based on a weighted average of the adversarial loss and the additional training loss.

16. The method of claim 15 , wherein respective weights assigned to the adversarial loss and the pixel-wise loss are adjusted during the training process to balance between a sharpness and a level of artifacts in an image produced by the second artificial neural network.

17. The method of claim 14 , wherein the ground truth MRI image is a fully sampled MRI image.

18. The method of claim 10 , wherein the first artificial neural network and the second artificial neural network are trained together in an end-to-end manner.

Assignees

• Shanghai United Imaging Intelligence Co Ltd

Inventors

• Chen Zhang
• Dong Siyuan
• Sun Shanhui
• Chen Xiao
• Liu Yikang
• Chen Terrence

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06N3/047 Primary

  Probabilistic or stochastic networks · CPC title

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title