Time compressed dynamic MR deep learning reconstruction
Systems and methods for reconstruction for a medical imaging system. Non-Cartesian k-space data is acquired using a dynamic MR sequence. A time compression network compresses the non-Cartesian data. The compressed data is used for reconstruction of an image. The time compression network is configured to reduce the (time and memory) complexity of the reconstruction process.
1 . A method of reconstruction for a medical imaging system, the method comprising:
scanning a patient by the medical imaging system, the scanning acquiring k-space scan data using a dynamic MR sequence that includes at least a time component;
applying, by a time compression network trained jointly with an unrolled iterative reconstruction network, a learned temporal compression to the k-space scan data to generate a time compression matrix that represents temporal correlations of the dynamic MR sequence;
reconstructing an image using the unrolled iterative reconstruction network that includes at least a data-consistency step, wherein the time compression matrix is used within the data-consistency step; and
outputting the image.
2 . The method of claim 1 , wherein the dynamic MR sequence comprises a GRASP (Golden-angle RAdial Sparse Parallel imaging) sequence.
3 . The method of claim 1 , further comprising:
applying an orthogonalization procedure at an end of the time compression network.
4 . The method of claim 3 , wherein the orthogonalization procedure comprises a Gram-Schmidt orthonormalization procedure, a Cayley transformation, or a Householder transformation.
5 . The method of claim 1 , further comprising:
applying a decompression matrix at an end of the unrolled iterative reconstruction network.
6 . The method of claim 1 , wherein the time compression network is first trained offline using supervised learning and ground truth images to generate target compression matrices; wherein the unrolled iterative reconstruction network is then trained end to end with the trained time compression network.
7 . The method of claim 1 , wherein an input of the time compression network is the k-space scan data comprising a three-dimensional matrix of size: readout size×number of coils×number of time points, or as a two-dimensional matrix of size: number of coils×number of time points.
8 . The method of claim 1 , wherein an output of the time compression network is a matrix of size: number of compressed time components×number of time points, wherein the number of compressed time components is predefined by an operator.
9 . The method of claim 8 , wherein the number of compressed time components is between five and ten and wherein the number of time points is greater than one hundred.
10 . The method of claim 1 , wherein the time compression network comprises multiple fully connected layers with nonlinear activation functions of transformer encoder layers.
11 . The method of claim 1 , wherein the time compression network is trained with sequence of variable time points.
12 . A system for time compressed dynamic magnetic resonance deep learning reconstruction, the system comprising:
a medical imaging system configured to acquire k-space scan data using a non-Cartesian dynamic sequence;
a time compression network configured to compress the k-space scan data; and
an unrolled iterative reconstruction network trained jointly with the unrolled iterative reconstruction network, the unrolled iterative reconstruction network configured to reconstruct an image, wherein the unrolled iterative reconstruction network is configured to apply, using the time compression network, a learned temporal compression to the k-space scan data to generate a time compression matrix that represents temporal correlations of the non-Cartesian dynamic sequence, wherein the unrolled iterative reconstruction network includes at least a data-consistency step, wherein the time compression matrix is used within the data-consistency step.
13 . The system of claim 12 , further comprising:
a display configured to display the image.
14 . The system of claim 12 , wherein the non-Cartesian dynamic sequence comprises a GRASP (Golden-angle RAdial Sparse Parallel imaging) sequence.
15 . The system of claim 12 , wherein the time compression network is first trained offline using supervised learning and ground truth images to generate target compression matrices; wherein the unrolled iterative reconstruction network is then trained end to end with the trained time compression network.
16 . The system of claim 12 , wherein an input of the time compression network is the k-space scan data comprising a three-dimensional matrix of size: readout size×number of coils×number of time points, or as a two-dimensional matrix of size: number of coils×number of time points and wherein an output of the time compression network is a matrix of size: number of compressed time components×number of time points, wherein the number of compressed time components is predefined by an operator.
17 . The system of claim 16 , wherein the number of compressed time components is between five and ten and wherein the number of time points is greater than one hundred.
18 . A non-transitory computer readable storage medium comprising a set of computer-readable instructions stored thereon which, when executed by at least one processor cause the processor to:
acquire non-Cartesian k-space scan data that includes at least a time component;
apply, by a time compression network trained jointly with an unrolled iterative reconstruction network, a learned temporal compression to the non-Cartesian k-space scan data to generate a time compression matrix that represents temporal correlations of the non-Cartesian k-space scan data;
reconstruct an image using an unrolled iterative reconstruction that includes at least a data-consistency step, wherein the time compression matrix is used within the data-consistency step; and
output the image.
19 . The non-transitory computer readable storage medium of claim 18 , wherein the non-Cartesian k-space scan data is acquired using a GRASP (Golden-angle RAdial Sparse Parallel imaging) sequence.
20 . The non-transitory computer readable storage medium of claim 18 , wherein an input of the time compression network is the non-Cartesian k-space scan data comprising a two-dimensional matrix of size: number of coils×number of time points and wherein the time compression matrix is a matrix of size: number of compressed time components×number of time points, wherein the number of compressed time components is predefined by an operator.