Generalized multi-channel MRI reconstruction using deep neural networks
A method for magnetic resonance imaging acquires multi-channel subsampled k-space data using multiple receiver coils; performs singular-value-decomposition on the multi-channel subsampled k-space data to produce compressed multi-channel k-space data which normalizes the multi-channel subsampled k-space data; applies a first center block of the compressed multi-channel k-space data as input to a first convolutional neural network to produce a first estimated k-space center block that includes estimates of k-space data missing from the first center block; generates an n-th estimated k-space block by repeatedly applying an (n−1)-th estimated k-space center block combined with an n-th center block of the compressed multi-channel k-space data as input to an n-th convolutional neural network to produce an n-th estimated k-space center block that includes estimates of k-space data missing from the n-th center block; reconstructs image-space data from the n-th estimated k-space block.
1. A method for magnetic resonance imaging using a magnetic resonance imaging apparatus, the method comprising
acquiring multi-channel subsampled k-space data using multiple receiver coils;
performing singular-value-decomposition on the multi-channel subsampled k-space data to produce compressed multi-channel k-space data which normalizes the multi-channel subsampled k-space data;
applying a first center block of the compressed multi-channel k-space data as input to a first convolutional neural network to produce a first estimated k-space center block that includes estimates of k-space data missing from the first center block;
generating an n-th estimated k-space block by repeating for each n from 2 to N>2, applying an (n−1)-th estimated k-space center block combined with an n-th center block of the compressed multi-channel k-space data as input to an n-th convolutional neural network to produce an n-th estimated k-space center block that includes estimates of k-space data missing from an n-th center block, wherein the (n−1)-th estimated k-space center block is smaller in size than the n-th center block;
reconstructing image-space data from the n-th estimated k-space block.
2. The method of claim 1 wherein generating an n-th estimated k-space block comprises, for each n from 2 to N>2, inserting into the produced n-th estimated k-space center block original data from the (n−1)-th estimated k-space center block combined with the n-th center block of the compressed multi-channel k-space data used as input.
3. The method of claim 1 wherein the n-th convolutional neural network is a residual network trained using center blocks extracted from fully sampled k-space datasets restrospectively subsampled with a fixed number of sampling masks.
4. The method of claim 1 wherein the multi-channel subsampled k-space data includes different time frames or echoes stacked along with channels.
5. The method of claim 1 wherein the k-space data are measured using a non-Cartesian trajectory and the data is first gridded onto a Cartesian grid.
6. The method of claim 1 wherein the multi-channel subsampled k-space data is in three-dimensional space.