CT super-resolution GAN constrained by the identical, residual and cycle learning ensemble (GAN-circle)

1. A system for generating a high resolution (HR) computed tomography (CT) image from a low resolution (LR) CT image, the system comprising:

a first generative adversarial network (GAN) comprising:

a first generative neural network (G) configured to receive a training LR image dataset and to generate a corresponding estimated HR image dataset, and

a first discriminative neural network (D Y ) configured to compare a training HR image dataset and the estimated HR image dataset;

a second GAN comprising:

a second generative neural network (F) configured to receive the training HR image dataset and to generate a corresponding estimated LR image dataset, and

a second discriminative neural network (D X ) configured to compare the training LR image dataset and the estimated LR image dataset; and

an optimization module configured to determine an optimization function based, at least in part, on at least one of the estimated HR image dataset and/or the estimated LR image dataset, the optimization function containing at least one loss function, the optimization module further configured to adjust a plurality of neural network parameters associated with at least one of the first GAN and/or the second GAN, to optimize the optimization function.

2. The system of claim 1 , wherein the at least one loss function is selected from the group comprising an adversarial loss function, a cyclic loss function, an identity loss function and a joint sparsifying transform loss function.

3. The system of claim 1 , wherein the at least one loss function comprises a cyclic loss function configured to implement a cycle consistency constraint.

4. The system of claim 1 , wherein the first generative neural network (G) and the second generative neural network (F) each comprise a respective feature extraction network and a respective reconstruction network, each feature extraction network comprising a plurality of skip connections.

5. The system of claim 1 , wherein the first discriminative neural network (D Y ) and the second discriminative neural network (D X ) each comprise a respective plurality of discriminator blocks coupled in series, each discriminator block comprising a convolutional stage, a bias stage, an instance norm stage, and a leaky rectified linear unit (ReLU) stage.

6. The system of claim 1 , wherein a trained first generative neural network (G) is configured to generate a HR CT image from an actual LR CT image, the actual LR CT image corresponding to low dose CT image data.

7. The system of claim 2 , wherein the group further comprises a supervision loss function.

8. A method for generating a high resolution (HR) computed tomography (CT) image from a low resolution (LR) CT image, the method comprising:

generating, by a first generative neural network (G), a corresponding estimated HR image dataset based, at least in part, on a received training LR image dataset;

comparing, by a first discriminative neural network (D Y ), a training HR image dataset and the estimated HR image dataset, the first generative neural network (G) and the first discriminative neural network (D Y ) included in a first generative adversarial network (GAN);

generating, by a second generative neural network (F), a corresponding estimated LR image dataset based, at least in part, on the training HR image dataset;

comparing, by a second discriminative neural network (D X ), the training LR image dataset and the estimated LR image dataset, the second generative neural network (F) and the second discriminative neural network (D X ) included in a second GAN;

determining, by an optimization module, an optimization function based, at least in part, on at least one of the estimated HR image dataset and/or the estimated LR image dataset, the optimization function containing at least one loss function; and

adjusting, by the optimization module, a plurality of neural network parameters associated with at least one of the first GAN and/or the second GAN, to optimize the optimization function.

9. The method of claim 8 , wherein the at least one loss function is selected from the group comprising an adversarial loss function, a cyclic loss function, an identity loss function and a joint sparsifying transform loss function.

10. The method of claim 8 , wherein the at least one loss function comprises a cyclic loss function configured to implement a cycle consistency constraint.

11. The method of claim 8 , wherein the first generative neural network (G) and the second generative neural network (F) each comprise a respective feature extraction network and a respective reconstruction network, each feature extraction network comprising a plurality of skip connections.

12. The method of claim 8 , wherein the first discriminative neural network (D Y ) and the second discriminative neural network (D X ) each comprise a respective plurality of discriminator blocks coupled in series, each discriminator block comprising a convolutional stage, a bias stage, an instance norm stage, and a leaky rectified linear unit (ReLU) stage.

13. The method of claim 8 , further comprising generating, by a trained first generative neural network (G), a HR CT image from an actual LR CT image, the actual LR CT image corresponding to low dose CT image data.

14. The method of claim 9 , wherein the group further comprises a supervision loss function.

15. A computer readable storage device having stored thereon instructions configured to generate a high resolution (HR) computed tomography (CT) image from a low resolution (LR) CT image, the instructions that when executed by one or more processors result in the following operations comprising:

generating, by a first generative neural network (G), a corresponding estimated HR image dataset based, at least in part, on a received training LR image dataset;

comparing, by a first discriminative neural network (D Y ), a training HR image dataset and the estimated HR image dataset, the first generative neural network (G) and the first discriminative neural network (D Y ) included in a first generative adversarial network (GAN);

generating, by a second generative neural network (F), a corresponding estimated LR image dataset based, at least in part, on the training HR image dataset;

comparing, by a second discriminative neural network (D X ), the training LR image dataset and the estimated LR image dataset, the second generative neural network (F) and the second discriminative neural network (D X ) included in a second GAN;

determining, by an optimization module, an optimization function based, at least in part, on at least one of the estimated HR image dataset and/or the estimated LR image dataset, the optimization function containing at least one loss function; and

adjusting, by the optimization module, a plurality of neural network parameters associated with at least one of the first GAN and/or the second GAN, to optimize the optimization function.

16. The device of claim 15 , wherein the at least one loss function is selected from the group comprising an adversarial loss function, a cyclic loss function, an identity loss function and a joint sparsifying transform loss function.

17. The device of claim 15 , wherein the at least one loss function comprises a cyclic loss function configured to implement a cycle consistency constraint.

18. The device of claim 15 , wherein the first generative neural network (G) and the second generative neural network (F) each comprise a respective feature extraction network and a respective reconstruction network, each feature extraction network comprising a plurality of skip connections.

19. The device of claim 15 , wherein the first discriminative neural network (D Y ) and the second discriminative neural network (D X ) each comprise a respective plurality of discriminator blocks coupled in series, each discriminator block comprising a convolutional stage, a bias stage, an instance norm stage, and a leaky rectified linear unit (ReLU) stage.

20. The device of claim 16 , wherein the group further comprises a supervision loss function.