Deep learning based methods to accelerate multi-spectral imaging
View Patent ↗A method for magnetic resonance imaging reconstructs images that have reduced under-sampling artifacts from highly accelerated multi-spectral imaging acquisitions. The method includes performing by a magnetic resonance imaging (MRI) apparatus an accelerated multi-spectral imaging (MSI) acquisition within a field of view of the MRI apparatus, where the sampling trajectories of different spectral bins in the acquisition are different; and reconstructing bin images using neural network priors learned from training data as regularization to reduce under-sampling artifacts.
1. A method for magnetic resonance imaging that reconstructs images having reduced under-sampling artifacts, the method comprising:
performing by a magnetic resonance imaging (MRI) apparatus an accelerated multi-spectral imaging (MSI) acquisition within a field of view of the MRI apparatus to produce under-sampled k-space data;
reconstructing from the under-sampled k-space data bin images using neural network priors learned from training data as regularization to reduce under-sampling artifacts; and
combining the reconstructed bin images to form a final image.
2. The method of claim 1 wherein an unrolled optimization algorithm is used to incorporate the neural network priors as regularization.
3. The method of claim 2 wherein the unrolled optimization algorithm is implemented using Iterative Shrinkage-Thresholding Algorithm (ISTA), alternating direction method of multipliers (ADMM), or a gradient descent algorithm.
4. The method of claim 1 wherein 3D convolutional neural networks (CNN) are used as the neural network priors.
5. The method of claim 4 wherein the 3D convolutional neural network (CNN) has convolutions along spatial and spectral dimensions.