Scan-specific recurrent neural network for image reconstruction

1. An image reconstruction system, comprising:

a memory configured to store a plurality of samples of biological, physiological, neurological or anatomical data with missing or corrupted samples and one or more deep learning models or neural networks, wherein the neural networks includes a recurrent neural network (RNN); and

a processor coupled to the memory and configured to:

obtain the plurality of samples,

determine missing or corrupted k-space samples using the plurality of samples that are under-sampled and the one or more deep learning models or neural networks,

wherein to determine the missing or corrupted k-space samples, the processor is further configured to iterate between performing one or more convolutions and applying a non-linear activation function a number of times to project or generate the missing or corrupted k-space data, wherein the number of times is pre-configured or user-inputted, and

reconstruct a magnetic resonance imaging (MRI) image using the plurality of samples and the determined missing or corrupted k-space.

2. The image reconstruction system of claim 1 , wherein the RNN applies the same network layer multiple times and uses at least one non-linear activation function.

3. The image reconstruction system of claim 2 , wherein to determine the missing or corrupted k-space samples the processor is configured to:

perform one or more convolutions on the plurality of samples to project the missing or corrupted k-space samples; and

apply the non-linear activation function while performing the one or more convolutions.

4. The image reconstruction system of claim 1 , wherein the processor is configured to train parameters of the one or more deep learning models or neural networks using a subset of k-space data that is not missing samples.

5. The image reconstruction system of claim 1 , wherein the processor is configured to:

generate the one or more deep learning models or neural networks, wherein the obtained plurality of samples are the only inputs used to generate the deep learning model or neural network.

6. The image reconstruction system of claim 1 , wherein the processor is configured to determine the missing or corrupted k-space samples to reconstruct the magnetic resonance image for a plurality of different patterns of under-sampled biological, physiological, neurological or anatomical data.

7. The image reconstruction system of claim 1 , further comprising:

a display configured to display the reconstructed MRI image.

8. The image reconstruction system of claim 1 , wherein the plurality of samples of biological, physiological, neurological or anatomical data with the missing or corrupted samples are obtained from a single MRI image of only a specific subject.

9. The image reconstruction system of claim 1 , wherein the plurality of samples are obtained in a Fourier transform domain and an inverse of the Fourier transform domain is performed to reconstruct the single MRI image, wherein a number of samples of the plurality of samples is less than a Nyquist rate.

10. An image reconstruction system, comprising:

a memory configured to store a plurality of samples of biological, physiological, neurological or anatomical data with missing or corrupted samples and one or more deep learning models or neural networks, wherein the neural networks includes a recurrent neural network (RNN); and

a processor coupled to the memory and configured to:

obtain the plurality of samples;

determine missing or corrupted k-space samples using the plurality of samples that are under-sampled and the one or more deep learning models or neural networks; and

reconstruct a magnetic resonance imaging (MRI) image using the plurality of samples and the determined missing or corrupted k-space,

wherein the RNN applies the same network layer multiple times and uses at least one non-linear activation function,

wherein to determine the missing or corrupted k-space samples the processor is configured to:

perform one or more convolutions on the plurality of samples to project the missing or corrupted k-space samples; and

apply the non-linear activation function while performing the one or more convolutions, and

wherein to determine the missing or corrupted k-space samples the processor is further configured to iterate between the performing the one or more convolutions and applying the non-linear activation function a number of times to project or generate missing or corrupted k-space data, wherein the number of times is pre-configured or user-inputted.

11. An image reconstruction system, comprising:

a memory configured to store a plurality of samples of biological, physiological, neurological or anatomical data with missing or corrupted samples and one or more deep learning models or neural networks, wherein the neural networks includes a recurrent neural network (RNN); and

a processor coupled to the memory and configured to:

train parameters of a first model and a second model using a subset of the k-space data that is complete,

determine a first set of reconstructed k-space data based on a plurality of samples and the first model,

determine a second set of reconstructed k-space data based on the plurality of samples and the second model, obtain the plurality of samples,

determine missing or corrupted k-space samples using the plurality of samples that are under-sampled and the one or more deep learning models or neural networks, and

reconstruct a magnetic resonance imaging (MRI) image using the plurality of samples and the determined missing or corrupted k-space wherein a reconstructed magnetic resonance imaging (MRI) image is further based on the first set and the second set of reconstructed k-space data,

wherein to determine the missing or corrupted k-space samples the processor is further configured to iterate between performing a one or more convolutions and applying a non-linear activation function a number of times to project or generate the missing or corrupted k-space data, wherein the number of times is pre-configured or user-inputted.

12. The image reconstruction system of claim 11 , wherein the subset of the k-space data that is complete does not have missing samples, wherein to determine the first set of reconstructed k-space data based on the plurality of samples and the first model the processor is configured to:

determine a first set of missing k-space data based on the first model and the plurality of samples; and

determine the first set of reconstructed k-space data further based on the first set of missing k-space data.

13. The image reconstruction system of claim 11 , wherein the processor is further configured to:

determine a first weight for the first set of reconstructed k-space data and a second weight for the second set of reconstructed k-space data; and

reconstruct the magnetic resonance imaging (MRI) image further based on weighted combination of the first and second set of reconstructed k-space data.

14. The image reconstruction system of claim 11 , further comprising:

a magnetic resonance imaging (MRI) scanner configured to obtain the plurality of samples of biological, physiological, neurological or anatomical data that include missing or corrupted k-space data.

15. The image reconstruction system of claim 11 , wherein the first model is a deep learning model or neural network that uses at least one non-linear activation function, wherein the deep learning model or the neural network is the recurrent neural network (RNN) that applies the same network layer multiple times.

16. The image reconstruction system of claim 15 , wherein the non-linear activation function is a rectified linear unit (ReLu) non-linear activation function.