Robust principal component analysis for separation of on and off-resonance in 3D multispectral MRI
View Patent ↗Accelerated 3D multispectral imaging (MSI) on a magnetic resonance imaging (MRI) system uses phase-encoding in two dimensions and frequency-encoding in the third. The method generates randomized k-space undersampling patterns that vary between different spectral bins to determine k-space samples to be acquired in each spectral bin; orders k-space samples into echo trains that determine gradient waveforms; initializes the gradient waveforms, RF waveforms, and timing information; plays the gradient and RF waveforms using the timing information to excite and refocus different spatial and spectral bin regions; acquires undersampled MRI signal data on the MRI system from the spatial and spectral bin regions; uses robust principal component analysis to reconstruct on-resonance images and off-resonance images represented as sets of low rank and sparse matrices from the undersampled MRI signal data; combines the on-resonance images and off-resonance images to yield a final image; and presents the final image on a display.
1. A method of accelerated 3D multispectral imaging (MSI) based on phase-encoding in two dimensions and frequency-encoding in the third, implemented on a magnetic resonance imaging (MRI) system, the method comprising:
generating randomized k-space undersampling patterns to determine k-space samples to be acquired in each spectral bin;
ordering k-space samples into echo trains that determine gradient waveforms;
initializing the gradient waveforms, RF waveforms, and timing information;
playing the gradient and RF waveforms using the timing information to excite and refocus different spatial and spectral bin regions;
acquiring undersampled MRI signal data on the MRI system from the spatial and spectral bin regions;
reconstructing on-resonance images and off-resonance images represented as sets of low rank and sparse matrices from the undersampled MRI signal data, wherein the reconstructing uses a robust-principal-component-analysis-based algorithm implemented on a CPU or GPU;
combining the on-resonance images and off-resonance images on the CPU or GPU to yield a final image;
presenting the final image on a display.
2. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises constraining a low rank matrix used in the robust-principal-component-analysis-based algorithm to have rank one.
3. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises representing the off-resonance images in a transform domain using a fixed, non-adaptive transform.
4. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises using CLEAR to impose an additional penalty to enable a combination with parallel imaging.
5. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises using a modified CLEAR penalty to impose an additional penalty to enable a combination of partial k-space acceleration and parallel imaging.
6. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises incorporating a constraint for autocalibrating parallel imaging.
7. The method of claim 6 , wherein incorporating the constraint for autocalibrating parallel imaging uses an explicit model of coil sensitivities.
8. The method of claim 6 , wherein incorporating the constraint for autocalibrating parallel imaging uses kernel encoding local correlations in k-space.
9. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises performing all matrix constructions on a coil-by-coil basis, yielding unique rank one and sparse matrices for each coil.
10. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises performing all matrix constructions at each readout location, yielding unique rank one and sparse matrices for each readout location and enabling reconstructions of each readout location to be carried out in parallel.
11. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises performing all matrix constructions within contiguous segments along the y dimension, yielding unique rank one and sparse matrices for each coil.
12. The method of claim 1 , wherein reconstructing on-resonance images and off-resonance images comprises incorporating a pre-computed on-resonance bin profile with a spatially-varying scaling to model the on-resonance signal, while the sparsity of the off-resonance signal is maintained.
13. The method of claim 1 , wherein generating randomized k-space undersampling patterns comprises using variable-density sampling patterns to decrease sampling density towards outer regions of k-space.
14. The method of claim 1 , wherein generating randomized k-space undersampling patterns comprises varying the undersampling patterns in different bins.
15. The method of claim 14 , wherein the variations in undersampling patterns in different bins are chosen to impose a separation between samples that is determined by the slice profiles and overlap.
16. The method of claim 1 , wherein generating randomized k-space undersampling patterns comprises using complementary Poisson-disc sampling patterns to generate the randomized k-space undersampling patterns and achieve a variation between sampling in different bins.
17. The method of claim 1 , wherein generating randomized k-space undersampling patterns comprises performing partial k-space sampling, where one side of k-space is not acquired and reconstructing images comprises an assumption of slow phase variation in the images.