Deep learning reconstruction of free breathing perfusion

1. A method for reducing artifacts in magnetic resonance imaging (MRI) data, the method comprising:

acquiring a k-space dataset of an anatomical subject using an MRI scanner;

using an iterative compressed sensing reconstruction method to generate a reconstructed image based on the k-space dataset, wherein the iterative compressed sensing reconstruction method uses (a) L1-norm based total variation constraints applied the temporal and spatial dimensions of the k-space dataset and (b) a low rank constraint; and

after the reconstructed image is generated, using a deep learning framework to generate an artifact image depicting motion artifacts present in the reconstructed image, wherein the deep learning framework comprises (a) a first deep learning network applied to magnitude parts of individual patches of the reconstructed image and (b) a second deep learning network applied to phase parts of the individual patches of the reconstructed image; and

subtracting the reconstructed image from the artifact image to yield a final image with the motion artifacts removed.

2. The method of claim 1 , wherein the k-space dataset is acquired using a radial simultaneous multi-slice (SMS) undersampled acquisition.

3. The method of claim 2 , wherein the radial SMS undersampled acquisition is performed using a plurality of k-space radial spokes with golden ratio-based angular spacing between individual spokes and spoke order.

4. The method of claim 1 , wherein the k-space dataset is acquired using a 3D acquisition.

5. The method of claim 1 , wherein the data is reconstructed using standard non-iterative techniques.

6. The method of claim 1 , wherein the k-space dataset is acquired using a spiral SMS undersampled acquisition.

7. The method of claim 1 , wherein the first deep learning network and the second deep learning network each comprises one or more Convolutional Neural Networks (CNNs).

8. The method of claim 7 , wherein the CNNs are trained using a plurality of fully sampled k-space datasets that are retrospectively under-sampled and reconstructed with L1 norm TV constraints.

9. The method of claim 7 , wherein the CNNs are trained to identify artifacts arising out of breathing motion.

10. The method of claim 7 , wherein the CNNs are trained to identify artifacts arising out of cardiac motion.

11. The method of claim 7 , wherein the CNNs are trained to identify artifacts arising out of breathing motion and cardiac motion.

12. The method of claim 1 , wherein the k-space dataset is acquired using an ECG-gated acquisition or an ungated acquisition that does not require an ECG signal.

13. A method for reducing artifacts in magnetic resonance imaging (MRI) data, the method comprising:

performing a radial SMS undersampled acquisition of a k-space dataset depicting anatomical and functional subject using a MRI scanner;

using an iterative compressed sensing reconstruction method to generate a reconstructed image based on the k-space dataset, wherein (a) each iteration of the iterative compressed sensing reconstruction method generates one or more estimated images and (b) the iterative compressed sensing reconstruction method uses a deep learning framework during each iteration to remove one or more motion artifacts from the estimated images,

wherein the deep learning framework comprises (a) a first deep learning network applied to magnitude parts of individual patches of the reconstructed image and (b) a second deep learning network applied to phase parts of the individual patches of the reconstructed image.

14. The method of claim 13 , wherein the iterative compressed sensing reconstruction method solves an objective function comprising (a) L1-norm based total variation constraints applied the temporal and spatial dimensions of the k-space dataset and (b) a low rank constraint.

15. The method of claim 13 , wherein the radial SMS undersampled acquisition is performed using a plurality of k-space radial spokes with golden ratio-based angular spacing between individual spokes.

16. The method of claim 13 , wherein the first deep learning network and the second deep learning network each comprises one or more Convolutional Neural Networks (CNNs).

17. The method of claim 16 , wherein the CNNs are trained using a plurality of fully sampled k-space datasets that are retrospectively under-sampled and reconstructed with L1 norm TV constraints.

18. The method of claim 16 , wherein the CNNs are trained to identify artifacts arising out of breathing motion.

19. The method of claim 16 , wherein the CNNs are trained to identify artifacts arising out of cardiac motion.

20. The method of claim 16 , wherein the CNNs are trained to identify artifacts arising out of breathing motion and cardiac motion.

21. A system for reducing artifacts in magnetic resonance imaging (MRI) data, the system comprising:

an MRI scanner configured to perform a radial SMS undersampled acquisition of a k-space dataset depicting anatomical subject

one or more computers configured to:

use an iterative compressed sensing reconstruction method to generate a reconstructed image based on the k-space dataset, wherein the iterative compressed sensing reconstruction method uses (a) L1-norm based total variation constraints applied to the temporal and spatial dimensions of the k-space dataset and (b) a low rank constraint;

after the reconstructed image is generated, using a deep learning framework to generate an artifact image depicting motion artifacts present in the reconstructed image, wherein the deep learning framework comprises (a) a first deep learning network applied to magnitude parts of individual patches of the reconstructed image and (b) a second deep learning network applied to phase parts of the individual patches of the reconstructed image; and

subtracting the reconstructed image from the artifact image to yield a final image with the motion artifacts removed.