Estimating diffusion metrics from diffusion- weighted magnetic resonance images using optimized k-q space sampling and deep learning
Diffusion metric maps are generated from a limited input of magnetic resonance data to a suitably trained machine learning algorithm, such as a suitably trained neural network. In general, a downsampling strategy is implemented in the joint k-q space to enable the simultaneous estimation of multiple different diffusion metrics from a more limited set of input diffusion-weighted images.
1. A method for generating diffusion metric maps from magnetic resonance data, the method comprising:
(a) accessing with a computer system, magnetic resonance data acquired from a subject, wherein the magnetic resonance data comprise:
first data that sample k-space with a first k-space sampling and sample q-space with a first q-space sampling;
second data that sample k-space with a second k-space sampling and sample q-space with a second q-space sampling;
wherein the first k-space sampling is higher than the second k-space sampling and the first q-space sampling is lower than the second q-space sampling;
(b) accessing a trained neural network with the computer system, wherein the trained neural network has been trained on training data in order to learn a mapping from magnetic resonance data to diffusion metrics corresponding to at least two different diffusion models; and
(c) inputting the magnetic resonance data to the trained neural network using the computer system, generating output as a plurality of diffusion metric maps, wherein the plurality of diffusion metric maps depict a spatial distribution in the subject of the diffusion metrics corresponding to the at least two different diffusion models.
2. The method of claim 1 , wherein the magnetic resonance data comprise jointly condensed k-q space data in which signal-to-noise tradeoffs are balanced by increasing k-space sampling for lower diffusion weighting b-values and increasing q-space sampling for lower spatial resolution acquisitions.
3. The method of claim 1 , wherein accessing the magnetic resonance data comprises acquiring the magnetic resonance data with a magnetic resonance imaging (MM) system and communicating the magnetic resonance data to the computer system.
4. The method of claim 1 , wherein the trained neural network comprises a convolutional neural network (CNN).
5. The method of claim 4 , wherein the CNN is a three-dimensional deep CNN.
6. The method of claim 5 , wherein the three-dimensional deep CNN is a three-dimensional deep residual network (ResNet) CNN.
7. The method of claim 1 , wherein the magnetic resonance data comprise diffusion-weighted magnetic resonance images reconstructed from the first data and the second data.
8. The method of claim 7 , wherein the diffusion-weighted magnetic resonance images are reconstructed from a combined dataset generated by combining the first data and the second data.
9. The method of claim 8 , wherein the combined dataset is generated by combining the first data and the second data using a spatially variant weighting of the first data and the second data.
10. The method of claim 1 , wherein the at least two different diffusion models are selected from the group consisting essentially of a diffusion tensor imaging model, a neurite orientation dispersion and density imaging model, a spherical mean technique micro-diffusion tensor model, and a spherical mean technique multi-compartment microstructural imaging model.
11. The method of claim 1 , wherein one of the at least two different diffusion models is a diffusion tensor imaging model and the diffusion metrics comprise at least one of fractional anisotropy, mean diffusivity, axial diffusivity, or radial diffusivity.
12. The method of claim 1 , wherein one of the at least two different diffusion models is a neurite orientation dispersion and density imaging model and the diffusion metrics comprise at least one of fraction of water in a restricted compartment, fraction of water in an isotropic compartment, orientation dispersion index, or hindered volume fraction.
13. The method of claim 1 , wherein one of the at least two different diffusion models is a spherical mean technique micro-diffusion tensor model and the diffusion metrics comprise at least one of fractional anisotropy, mean diffusivity, or radial diffusivity.
14. The method of claim 1 , wherein one of the at least two different diffusion models is a spherical mean technique multi-compartment microstructural imaging model and the diffusion metrics comprise at least one of intra-axonal water fraction, extra-neurite microscopic mean diffusivity, extra-neurite transverse microscopic diffusivity.
15. A method for generating diffusion metric maps from magnetic resonance data acquired with a magnetic resonance imaging (MM) system, the method comprising:
(a) acquiring diffusion-weighted magnetic resonance data from a subject with the MRI system, wherein the diffusion-weighted magnetic resonance data jointly sample k-q space such that k-space sampling is increased at lower diffusion weighting b-values and q-space sampling is increased at lower spatial resolutions;
(b) reconstructing diffusion-weighted images from the diffusion-weighted magnetic resonance data using a computer system;
(c) accessing a trained neural network with the computer system, wherein the trained neural network has been trained on training data in order to learn a mapping from diffusion-weighted images to diffusion metrics corresponding to at least two different diffusion models; and
(d) inputting the diffusion-weighted images to the trained neural network using the computer system, generating output as a plurality of diffusion metric maps, wherein the plurality of diffusion metric maps depict a spatial distribution in the subject of the diffusion metrics corresponding to the at least two different diffusion models.
16. The method of claim 15 , wherein the diffusion-weighted images consist of a non-diffusion weighted image acquired with a zero diffusion weighting b-value and nine diffusion-weighted images acquired with a non-zero diffusion weighting b-value.
17. The method of claim 16 , wherein the nine diffusion-weighted images include three images acquired at each of three different non-zero diffusion weighting b-values.
18. The method of claim 15 , wherein the diffusion-weighted magnetic resonance data are acquired over a range of diffusion weighting b-values spanning 0 mm/s 2 to 8,000 mm/s 2 .
19. The method of claim 15 , wherein the at least two different diffusion models are selected from the group consisting essentially of a diffusion tensor imaging model, a neurite orientation dispersion and density imaging model, a spherical mean technique micro-diffusion tensor model, and a spherical mean technique multi-compartment microstructural imaging model.