Deep Gray Matter (DGM) Segmentation using 3D Convolutional Neural Network: application to QSM
This work is based on:
- Jose Dolz, Christian Desrosiers, Ismail Ben Ayed, 3D fully convolutional networks for subcortical segmentation in MRI: A large-scale study, In NeuroImage, 2017
- joseabernal's solution for iSeg2017. Github
Current outcome
Accepted by ISMRM Workshop on Machine Learning 2018.
Some preliminary reports can be found at Medium (Part 1) (Part 2)
Highlight
Larger kernel size (7, 7, 3), add Batch Normalization and auxiliary feature input of spatial coordinates information.
Add wrapper for segmentation (inference).
How to use it (for training)
- Put QSM images in datasets/QSM/
- Put spatial coordinates maps in datasets/X/, datasets/Y/, datasets/Z/
- Put segmented ROI labels in datasets/label/
- Run segDGM_3DCNN.ipynb
How to use it (for segmenting nifti)
Example: python3 segDGM_3DCNN.py -i input_filename.nii.gz -o output_label.nii.gz
It uses pre-calculated weights in models/weights_optimal.h5