Fat suppression using neural networks
In a method for determining a fat-reduced MR image, a first MR image is provided having, apart from the other tissue constituents, MR signals from only one of the two fat constituents, the first MR image is applied to a trained ANN, which was trained by first MR training data as the input data, the training data including, apart from the other tissue constituents, MR signals from only the one of the two fat constituents, and using second MR training data as a base knowledge, the second MR training data including, apart from the other tissue constituents, no MR signals from the two fat constituents; and an MR output image is determined from the trained ANN, to which the first MR image was applied, as a fat-reduced MR image, wherein the fat-reduced MR image includes, apart from the other tissue constituents, no MR signals from the two fat constituents.
1. A computer-implemented method, performed using one or more processors and a non-transitory computer-readable storage medium, for determining a fat-reduced magnetic resonance (MR) image of a tissue under examination that has, apart from other tissue constituents, two fat constituents of different resonance frequency, the method comprising:
providing a first MR image, which has, apart from the other tissue constituents, MR signals from only one of the two fat constituents;
applying the first MR image to a trained artificial neural network, which was trained by first MR training data as input data, the training data including, apart from the other tissue constituents, MR signals from only the one of the two fat constituents, and using second MR training data as a base knowledge, the second MR training data including, apart from the other tissue constituents, no MR signals from the two fat constituents; and
determining an MR output image from the trained artificial neural network, to which the first MR image was applied, as a fat-reduced MR image, wherein the fat-reduced MR image includes, apart from the other tissue constituents, no MR signals from the two fat constituents.
2. The method as claimed in claim 1 , wherein the second MR training data has been generated by an imaging sequence in which a saturation inversion pulse suppresses the two fat constituents during signal acquisition.
3. The method as claimed in claim 1 , wherein the first MR image and the first MR training data were generated by an imaging sequence in which the one of the two fat constituents was suppressed by spectral fat saturation, while another of the two fat constituents is still contained in the respective MR signals which were used to generate the first MR image and the first MR training data.
4. The method as claimed in claim 1 , wherein the trained artificial neural network is a convolutional U-shaped neural network.
5. The method as claimed in claim 1 , wherein the first MR image and the first and second MR training data were acquired by an echoplanar technique.
6. The method as claimed in claim 1 , further comprising: determining a mask configured to select a subregion of the first MR image that contains image points having MR signals from only the one of the two fat constituents; and
applying the mask to the first MR image to select the subregion of the first MR image, wherein applying the first MR image to the trained artificial neural network comprises applying only the selected subregion selected by the mask to the trained artificial neural network.
7. The method as claimed in claim 1 , wherein the first MR training data and the second MR training data have been acquired using different diffusion encodings.
8. The method as claimed in claim 7 , wherein the first MR image has a first diffusion encoding, wherein the method further comprises applying a second MR image to the artificial neural network, the second MR image having a same slice position as the first MR image and having, apart from the other tissue constituents, no MR signals from the two fat constituents, wherein the second MR image includes a diffusion encoding that differs from the first diffusion encoding.
9. The method as claimed in claim 8 , further comprising applying a third MR image to the artificial neural network, the third MR image having: the same slice position as the first MR image, apart from the other tissue constituents, only the one of the two fat constituents, and a diffusion encoding that differs from the first diffusion encoding.
10. The method as claimed in claim 1 , wherein the first MR image comprises a first contrast and a first resolution, and wherein the method further comprises applying a second MR image to the artificial neural network, the second MR image having a low resolution compared with the first MR image and a second contrast that differs from the first contrast, wherein the MR output image includes the first resolution and the second contrast.
11. A computer program product embodied on the non-transitory computer-readable storage medium and which is directly loadable into a memory of a magnetic resonance device, when executed by the one or more processors of the magnetic resonance device, causes the magnetic resonance device to perform the method as claimed in claim 1 .
12. A non-transitory computer-readable storage medium with an executable program stored thereon, that when executed, instructs a processor to perform:
providing a first MR image, which has, apart from the other tissue constituents, MR signals from only one of the two fat constituents;
applying the first MR image to a trained artificial neural network, which was trained by first MR training data as input data, the training data including, apart from the other tissue constituents, MR signals from only the one of the two fat constituents, and using second MR training data as a base knowledge, the second MR training data including, apart from the other tissue constituents, no MR signals from the two fat constituents; and
determining an MR output image from the trained artificial neural network, to which the first MR image was applied, as a fat-reduced MR image, wherein the fat-reduced MR image includes, apart from the other tissue constituents, no MR signals from the two fat constituents.
13. An apparatus for determining a fat-reduced magnetic resonance (MR) image of a tissue under examination that has, apart from other tissue constituents, two fat constituents of different resonance frequency, the apparatus comprising:
an artificial trained neural network; and
a processing device configured to:
provide a first MR image including, apart from the other tissue constituents, MR signals from only one of the two fat constituents;
apply the first MR image to the trained artificial neural network, wherein the artificial trained neural network has been trained by first MR training data as input data, the first MR training data including, apart from the other tissue constituents, MR signals from only the one of the two fat constituents, and using second MR training data as a base knowledge, the second MR training data including, apart from the other tissue constituents, no MR signals from the two fat constituents; and
determine an MR output image, from the trained artificial neural network to which the first MR image was applied, as a fat-reduced MR image that includes, apart from the other tissue constituents, no MR signals from the two fat constituents.
14. The apparatus as claimed in claim 13 , wherein the second MR training data has been generated by an imaging sequence in which a saturation inversion pulse suppresses the two fat constituents during signal acquisition.
15. The apparatus as claimed in claim 13 , wherein the first MR image and the first MR training data were generated by an imaging sequence in which the one of the two fat constituents was suppressed by spectral fat saturation, while the other of the two fat constituents is still contained in the respective MR signals which were used to generate the first MR image and the first MR training data.
16. The apparatus as claimed in claim 13 , wherein the trained artificial neural network is a convolutional U-shaped neural network.
17. The apparatus as claimed in claim 13 , wherein the first MR image and the first and second MR training data were acquired by the echoplanar technique.
18. The apparatus as claimed in claim 13 , wherein the processing device is further configured to:
determine a mask configured to select a subregion of the first MR image that contains image points having MR signals from only the one of the two fat constituents; and
apply the mask to the first MR image to select the subregion of the first MR image, wherein applying the first MR image to the trained artificial neural network comprises applying only the selected subregion selected by the mask to the trained artificial neural network.
19. The apparatus as claimed in claim 13 , wherein the first MR training data and the second MR training data have been acquired using different diffusion encodings.
20. The apparatus as claimed in claim 19 , wherein:
the first MR image has a first diffusion encoding; and
the processing device is configured to:
apply a second MR image to the artificial neural network, the second MR image having a same slice position as the first MR image and having, apart from the other tissue constituents, no MR signals from the two fat constituents, the second MR image including a diffusion encoding that differs from the first diffusion encoding; and
apply a third MR image to the artificial neural network, the third MR image having: the same slice position as the first MR image, apart from the other tissue constituents, only the one of the two fat constituents, and a diffusion encoding that differs from the first diffusion encoding.