Super-resolution anatomical magnetic resonance imaging using deep learning for cerebral cortex segmentation
Super-resolution images are generated from standard-resolution images acquired with a magnetic resonance imaging (“MRI”) system. More particularly, super-resolution (e.g., sub-millimeter isotropic resolution) images are generated from standard-resolution images (e.g., images with 1 mm or coarser isotropic resolution) using a deep learning algorithm, from which accurate cortical surface reconstructions can be generated.
1. A method for generating a cortical surface reconstruction from standard-resolution magnetic resonance images, the method comprising:
(a) accessing standard-resolution magnetic resonance image data with a computer system, wherein the standard-resolution magnetic resonance image data depict a brain of a subject;
(b) generating upsampled image data with the computer system by upsampling the standard-resolution magnetic resonance image data to a target spatial resolution that is finer than the standard resolution;
(c) accessing a neural network with the computer system, wherein the neural network has been trained on training data to learn a mapping from upsampled image data to native high-resolution image data, the training data including
first image data having a first spatial resolution representative of a native high resolution,
second image data having a second spatial resolution that is coarser than the first spatial resolution, wherein the second image data are generated by downsampling the first image data, and
third image data having a third spatial resolution representative of an upsampled high resolution, wherein the third image data are generated by upsampling the second image data;
(d) inputting the upsampled image data to the neural network using the computer system, generating output as residual image data that represent image differences relative to the upsampled image data; and
(e) generating super-resolution image data by combining the upsampled image data and the residual image data.
2. The method of claim 1 , wherein the standard-resolution magnetic resonance image data have a spatial resolution coarser than or equal to 1 millimeter isotropic.
3. The method of claim 1 , wherein the super-resolution image data have a spatial resolution finer than 1 millimeter isotropic.
4. The method of claim 1 , wherein the neural network is a convolutional neural network.
5. The method of claim 4 , wherein the convolutional neural network implements deep learning.
6. The method of claim 4 , wherein the convolutional neural network is a very-deep super-resolution (VDSR) neural network.
7. The method of claim 1 , wherein the upsampled image data are generated using a cubic spline interpolation.
8. The method of claim 1 , further comprising generating a cortical surface reconstruction from the super-resolution image data, wherein the cortical surface reconstruction represents a three-dimensional representation of a cortical surface of the brain of the subject.
9. The method of claim 8 , further comprising computing cortical thickness values for the brain of the subject based on the cortical surface reconstruction.
10. The method of claim 1 , wherein the neural network is a pre-trained neural network and step (c) comprises fine-tuning the pre-trained neural network using additional training data.
11. The method of claim 1 , wherein the neural network is configured to generate super-resolution magnetic resonance images from lower resolution magnetic resonance images.
12. The method of claim 1 , wherein the neural network is trained using the training data by minimizing a loss function to learn a mapping between the native high resolution and the upsampled high resolution.