Deep learning-based water-fat separation from dual-echo chemical shift encoded imaging
View Patent ↗A method for magnetic resonance imaging performs chemical shift encoded imaging to produce complex dual-echo images which are then applied (with imaging parameters) as input to a deep neural network to produce as output water-only and fat-only images. The deep neural network can be trained with ground truth water/fat images derived from chemical shift encoded images using a conventional water-fat separation algorithm such as projected power approach, IDEAL, or VARPRO. The chemical shift encoded imaging comprises performing an image acquisition with the MRI scanner via a spoiled-gradient echo sequence or a spin-echo sequence.
1. A method for magnetic resonance imaging, comprising:
performing by an MRI scanner chemical shift encoded imaging to acquire complex dual-echo images;
applying by the MRI scanner the complex dual-echo images and imaging parameters as input to a deep neural network to produce as output separate water-only and fat-only images;
storing and displaying by the MRI scanner the separate water-only and fat-only images for diagnostic or therapeutic purposes;
wherein the deep neural network is trained with ground truth water/fat images derived from chemical shift encoded images using a conventional water-fat separation algorithm;
wherein performing chemical shift encoded imaging comprises performing an image acquisition with the MRI scanner via a spoiled-gradient echo sequence or a spin-echo sequence.
2. The method of claim 1 wherein performing chemical shift encoded imaging comprises using undersampling patterns selected from the group consisting of Cartesian variable density Poisson disc sampling, cones acquisition, and radial acquisition.
3. The method of claim 1 wherein performing chemical shift encoded imaging comprises reconstructing the complex dual-echo images using parallel imaging and/or compressed sensing reconstruction approaches.
4. The method of claim 1 wherein the input to the deep neural network comprises both phase and magnitude of the complex dual-echo images.
5. The method of claim 1 wherein the imaging parameters include imaging parameters for water-fat separation that comprise TEs of dual-echo images.
6. The method of claim 1 wherein the deep neural network is trained using a loss function selected from conventional l 1 , RMSE (root-of-mean-squared error), a mixed l 1 -SSIM loss, perceptual loss, or other loss function in which physical models are integrated.
7. The method of claim 1 wherein the deep neural network comprises two deep neural networks that output the separate water and fat images.
8. The method of claim 1 wherein the deep neural network comprises a single deep neural network that produces both water and fat images as the outputs.
9. The method of claim 1 wherein the deep neural network is a modified U-Net that has a hierarchical network architecture with global shortcuts and densely connected local shortcuts; wherein at each hierarchical level, there are several convolutional blocks; wherein image features are extracted using 3×3 convolutional kernels, followed by a Parametric Rectified Linear Unit (PReLU).