applyDiffNet_unet.py
simple implementation of pretained convolutional neural networks which reconstruct fractional anisotrpy (FA) and mean diffusivity (MD) maps from diffusion tensor imaging (DTI) scans.
This code reads in a dti image set in .nii.gz format. These dicom files should contain DWI with b=1000s/mm2 and 3, 6, or 20 diffusion encoding directions plus one b=0 reference image. This code assumes that data contains b=0 images first followed by all DWI.
Data should be formatted as: [Nx,Ny,Nslice,Ndir]
** these networks are trained on data with b=1000s/mm2. This will not work properly for other b-values **
Code outputs dFA and dMD maps (i.e. FA and MD estimates) in matlab (.mat) and numpy (.npy) formats.
Requirements: python with the following libraries
numpy
scipy
nibabel - only needed to load .nii.gz images
matplotlib - only needed to display output
keras - https://keras.io/
Usage:
python <pathToCode>\applyDiffNet_unet.py <NeuralNetLoc> <ImageLoc> <OutLoc>
where: NeuralNetLoc: path to folder containing .h5 keras neural network files (provided)
ImageLoc: path to image file (.nii.gz)
OutLoc: path for desired output
Example image files are provided with 3, 6, and 20 diffusion encoding directions. After successful completion, you should see one slice of the reconstructed dFA and dMD maps in a dialog box:
Note: I recently noticed some weird behavior with older versions of keras/tensorflow in which the models would load and compute predictions, but with highly innacurate FA/MD values. The models appear to be working correctly with the following versions of keras/tensorflow:
keras 2.4.3
tensorflow 2.2.0
Keep this in mind if the resulting maps do not resemble the above images.
Eric Aliotta, PhD
University of Virginia, Department of Radiation Oncology
07.10.2020
eric.aliotta@virginia.edu