Noise suppression for wave-CAIPI

Patent metadata

Field	Value
Publication number	US-10895622-B2
Application number	US-201916289985-A
Country	US
Kind code	B2
Filing date	Mar 1, 2019
Priority date	Mar 13, 2018
Publication date	Jan 19, 2021
Grant date	Jan 19, 2021

Abstract

Techniques are disclosed to leverage the use of neural networks or similar machine learning algorithms to de-noise highly accelerated Wave-CAIPIRINHA scans. The described techniques facilitate the generation of 3D sequences using a greatly reduced scan time, with the resulting images having a high spatial resolution and an improved SNR compared to conventional approaches.

First claim

What is claimed is:

1. A method, comprising:
accessing, via one or more processors, an under-sampled dataset included in a Wave dataset, the under-sampled dataset being associated with a magnetic resonance imaging (MRI) scan, the Wave dataset corresponding to two added oscillating wave gradients thereby yielding a Wave dataset having a corkscrew trajectory in k-space;
generating, via the one or more processors, a reconstructed dataset from the under-sampled dataset;
training, via the one or more processors, a convolutional neural network by calculating filter weights that result in the convolutional neural network providing an output image that substantially matches a difference image;
propagating, via the one or more processors, the reconstructed dataset through the trained convolutional neural network to generate one or more de-noised output images; and
displaying, via the one or more processors, the one or more de-noised output images.

2. The method of claim 1 , wherein the under-sampled dataset includes a Wave-controlled aliasing in volumetric parallel imaging (CAIPIRINHA) dataset.

3. The method of claim 2 , wherein the Wave dataset further includes at least one fully-sampled dataset that is equivalent to a ground truth.

4. The method of claim 3 , wherein the difference image is a result of a subtraction of the ground truth and an accelerated MRI scan that produces the under-sampled dataset.

5. The method of claim 1 , wherein the act of generating the reconstructed under-sampled dataset includes the use of a Wave-controlled aliasing in parallel imaging (CAIPI) reconstruction.

6. The method of claim 5 , wherein the act of generating the reconstructed under-sampled dataset includes the use of a gradient calibration.

7. The method of claim 1 , further comprising:
simulating, via the one or more processors, at least a portion of the Wave dataset by convolving a fully-sampled No-Wave scan with a Wave Point-Spread-Function (PSF) and applying retrospective under-sampling.

8. The method of claim 4 , wherein the difference image resulting from the subtraction of the ground truth and the accelerated MRI scan yields only white noise.

9. The method of claim 4 , wherein the accelerated scans include the use of retrospectively under-sampling the at least one fully-sampled scan without the use of image registration.

10. The method of claim 5 , wherein the act of propagating the reconstructed dataset through the trained convolutional neural network to generate one or more de-noised output images is part of a residual learning procedure.

11. A convolutional neural network, comprising:
a storage medium configured to store an under-sampled dataset included in a Wave dataset, the under-sampled dataset being associated with a magnetic resonance imaging (MRI) scan, the Wave dataset corresponding to two added oscillating wave gradients thereby yielding a Wave dataset having a corkscrew trajectory in k-space; and
one or more processors configured to:
generate a reconstructed dataset from the under-sampled dataset;
train a convolutional neural network by calculating filter weights that result in the convolutional neural network providing an output image that substantially matches a difference image;
propagate the reconstructed dataset through the trained convolutional neural network to generate one or more de-noised output images; and
display the one or more de-noised output images.

12. The convolutional neural network of claim 11 , wherein the under-sampled dataset includes a Wave-controlled aliasing in volumetric parallel imaging (CAIPIRINHA) dataset.

13. The convolutional neural network of claim 12 , wherein the Wave dataset further includes at least one fully-sampled dataset that is equivalent to a ground truth.

14. The convolutional neural network of 10 , wherein the difference image is a result of a subtraction of the ground truth and an accelerated MRI scan that produces the under-sampled dataset.

15. The convolutional neural network of claim 11 , wherein the one or more processors are further configured to generate the reconstructed under-sampled dataset using a Wave-controlled aliasing in parallel imaging (CAIPI) reconstruction.

16. The convolutional neural network of claim 15 , wherein the one or more processors are further configured to generate the reconstructed under-sampled dataset using a gradient calibration.

17. The convolutional neural network of claim 11 , wherein the one or more processors are further configured to simulate at least a portion of the Wave dataset by convolving a fully-sampled No-Wave scan with a Wave Point-Spread-Function (PSF) and applying retrospective under-sampling.

18. The convolutional neural network of claim 11 , wherein the convolutional neural network is associated with a U-NET architecture.

19. A non-transitory computer-readable medium associated with a convolutional neural network, the non-transitory medium having instructions stored thereon that, when executed by one or more processors, cause the one or more processors to:
access an under-sampled dataset included in a Wave dataset, the under-sampled dataset being associated with a magnetic resonance imaging (MRI) scan, the Wave dataset corresponding to two added oscillating wave gradients thereby yielding a Wave dataset having a corkscrew trajectory in k-space;
generate a reconstructed dataset from the under-sampled dataset;
train the convolutional neural network by calculating filter weights that result in the convolutional neural network providing an output image that substantially matches a difference image;
propagate the reconstructed dataset through the trained convolutional neural network to generate one or more de-noised output images; and
display the one or more de-noised output images.

20. The non-transitory computer-readable medium of claim 19 , wherein:
the under-sampled dataset includes a Wave-controlled aliasing in volumetric parallel imaging (CAIPIRINHA) dataset,
the Wave dataset further includes at least one fully-sampled dataset that is equivalent to a ground truth, and
the difference image is a result of a subtraction of the ground truth and an accelerated MRI scan that produces the under-sampled dataset.

21. The non-transitory computer-readable medium of claim 19 , further storing instructions that, when executed by one or more processors, cause the one or more processors to generate the reconstructed under-sampled dataset using a Wave-controlled aliasing in parallel imaging (CAIPI) reconstruction.

22. The non-transitory computer-readable medium of claim 21 , further storing instructions that, when executed by one or more processors, cause the one or more processors to generate the reconstructed under-sampled dataset using a gradient calibration.

23. The non-transitory computer-readable medium of claim 19 , further storing instructions that, when executed by one or more processors, cause the one or more processors to simulate at least a portion of the Wave dataset by convolving a fully-sampled No-Wave scan with a Wave Point-Spread-Function (PSF) and applying retrospective under-sampling.

Assignees

• Siemens Healthcare Gmbh
• Massachusetts Gen Hospital

Inventors

• Polak Daniel
• Raithel Esther
• Setsompop Kawin

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/08

  Learning methods · CPC title