MRI reconstruction using deep learning, generative adversarial network and acquisition signal model
A method for diagnostic imaging includes measuring undersampled data y with a diagnostic imaging apparatus; linearly transforming the undersampled data y to obtain an initial image estimate {tilde over (x)}; applying the initial image estimate {tilde over (x)} as input to a generator network to obtain an aliasing artifact-reduced image x̆ as output of the generator network, where the aliasing artifact-reduced image x̆ is a projection onto a manifold of realistic images of the initial image estimate {tilde over (x)}; and performing an acquisition signal model projection of the aliasing artifact-reduced x̆ onto a space of consistent images to obtain a reconstructed image {circumflex over (x)} having suppressed image artifacts.
1. A method for diagnostic imaging comprising:
measuring undersampled data y with a diagnostic imaging apparatus;
linearly transforming the undersampled data y to obtain an initial image estimate {tilde over (x)};
applying the initial image estimate {tilde over (x)} as input to a generator network to obtain an aliasing artifact-reduced image x̆ as output of the generator network,
wherein the aliasing artifact-reduced image x̆ is a projection onto a manifold of realistic images of the initial image estimate {tilde over (x)};
performing an acquisition signal model projection of the aliasing artifact-reduced image x̆ onto a space of consistent images to obtain a reconstructed image {circumflex over (x)} having suppressed image artifacts.
2. The method of claim 1 wherein the diagnostic imaging apparatus is an MRI scanner, and wherein the undersampled data is k-space data.
3. The method of claim 1 wherein linearly transforming the undersampled data y comprises zero padding the undersampled data y.
4. The method of claim 1 wherein linearly transforming the undersampled data y comprises finding an approximate zero-filling reconstruction from the undersampled data y.
5. The method of claim 1 wherein the generator network is trained to learn the projection onto the manifold of realistic images using a set of training images X and corresponding set of undersampled measurements Y using least-squares generative adversarial network techniques in tandem with a discriminator network to learn texture details and supervised cost function to control high-frequency noise.
6. The method of claim 5 wherein the supervised cost function comprises a mixture of smooth l 2 cost and non-smooth l 1 cost.
7. The method of claim 5 wherein the discriminator network is a multilayer deep convolutional neural network.
8. The method of claim 5 wherein the discriminator network is trained using a least squares cost for a discriminator decision.
9. The method of claim 1 wherein the generator network is a deep residual network with skip connections.
10. The method of claim 1 wherein performing an acquisition signal model projection is implemented as part of the generator network using a soft least-squares penalty during training of the generator network.
11. The method of claim 1 wherein the reconstructed image {circumflex over (x)} is applied to the generator network to obtain a second aliasing artifact-reduced image, and the second aliasing artifact-reduced image is projected onto the space of consistent images to obtain a final reconstructed image.