Model-based reconstruction for grasp MRI
Systems and methods for a deep learning reconstruction network with computationally light and efficient CNN architecture and a training strategy tailored to image reconstruction of dynamic multi-coil GRASP MRI. The configuration of the size of the network used in training time may be adjusted, which allows for higher accelerations and different hardware constraints.
1 . A method for reconstructing non-cartesian magnetic resonance imaging (MRI) data, the method comprising:
training a convolutional neural network (CNN) on image pairs to minimize a loss;
selecting a number of iterations for a reconstruction network;
assembling the reconstruction network with the number of iterations of the CNN, each iteration further including a data consistency step, wherein the CNN of different iterations share the same weights;
training the reconstruction network end to end; and
applying the reconstruction network to non-cartesian MRI data of a patient acquired from a medical imaging procedure.
2 . The method of claim 1 , wherein the non-cartesian MRI data is acquired using Golden-angle RAdial Sparse Parallel (GRASP).
3 . The method of claim 1 , wherein the CNN comprises three deformable 2D UNet branches for xt-, yt-, and xy-domains with weight sharing between all of the three deformable 2D UNet branches.
4 . The method of claim 1 , wherein the CNN is configured to perform regularization.
5 . The method of claim 1 , wherein the CNN is trained using training data comprising image pairs including initial NUFFT reconstructions and corresponding ground truth.
6 . The method of claim 1 , wherein the data consistency step comprises conjugate gradient data consistency.
7 . The method of claim 1 , wherein selecting the number of iterations is based on available hardware resources for training the reconstruction network.
8 . The method of claim 1 , wherein selecting the number of iterations is based on a targeted acceleration.
9 . The method of claim 1 , wherein selecting the number of iterations is based on a clinical application.
10 . A system for magnetic resonance imaging (MRI) reconstruction, the system comprising:
an MR imaging device configured to acquire non cartesian MRI data of a patient;
a reconstruction network configured to input the non cartesian MRI data and output a representation, the reconstruction network comprising a selected number of iterations, wherein each iteration includes a convolutional neural network (CNN) and a data consistency step, wherein the reconstruction network is trained in two stages, wherein in a first stage of the two stages a single iteration of the CNN is trained, wherein in a second stage of the two stages, the reconstruction network comprising the selected number of iterations is trained end to end wherein weights for the CNN in each iteration are initialized with weights learned in the first stage; and
a display configured to display the representation.
11 . The system of claim 10 , wherein the non cartesian MRI data is acquired using radial sampling.
12 . The system of claim 10 , wherein the CNN comprises three deformable 2D UNet branches for xt-, yt-, and xy-domains with weight sharing between all of the three deformable 2D UNet branches.
13 . The system of claim 10 , wherein the number of iterations is selected based on at least one of available hardware resources for training the reconstruction network, a targeted acceleration, or a clinical application for the representation.
14 . The system of claim 10 , wherein the data consistency step comprises conjugate gradient data consistency.
15 . The system of claim 10 , wherein in the first stage, the CNN is trained using image pairs including initial NUFFT reconstructions and corresponding ground truth.
16 . The system of claim 10 , wherein the CNN is configured to perform regularization.
17 . A method for MRI reconstruction, the method comprising:
acquiring non cartesian MRI data of a patient;
reconstructing a representation of the patient using a reconstruction network configured to input the non cartesian MRI data and output a representation, the reconstruction network comprising a selected number of iterations, wherein each iteration includes a convolutional neural network (CNN) and a data consistency step, wherein the reconstruction network is trained in two stages, wherein in a first stage of the two stages a single iteration of the CNN is trained, wherein in a second stage of the two stages, the reconstruction network comprising the selected number of iterations is trained end to end wherein weights for the CNN in each iteration are initialized with weights learned in the first stage; and
displaying the representation of the patient.
18 . The method of claim 17 , wherein the number of iterations is selected based on at least one of available hardware resources for training the reconstruction network, a targeted acceleration, or a clinical application for the representation.
19 . The method of claim 17 , wherein the CNN comprises three deformable 2D UNet branches for xt-, yt-, and xy-domains with weight sharing between all of the three deformable 2D UNet branches.
20 . The method of claim 17 , wherein the data consistency step comprises conjugate gradient data consistency.