Unsupervised Learning with CNNs for Image Registration
This repository incorporates several variants, first presented at CVPR2018 (initial unsupervised learning) and then MICCAI2018 (probabilistic & diffeomorphic formulation)
keywords: machine learning, convolutional neural networks, alignment, mapping, registration
Due to popular demand, we added a preliminary pytorch version; please see the pytorch folder.
Note: we will soon be updating a new voxelmorph version that will integrate both
New core tutorial provides intuition to voxelmorph!
See our learning method to automatically build atlases using VoxelMorph, to appear at NeurIPS2019.
It might be useful to have each folder inside the
ext folder on your python path.
assuming voxelmorph is setup at
If you would like to train/test your own model, you will likely need to write some of the data loading code in 'datagenerator.py' for your own datasets and data formats. There are several hard-coded elements related to data preprocessing and format.
These instructions are for the MICCAI2018 variant using
If you'd like to run the CVPR version (no diffeomorphism or uncertainty measures, and using CC/MSE as a loss function) use
train_miccai2018.pyto the location of your image files.
train_miccai2018.pywith options described in the main function at the bottom of the file. Example:
train_miccai2018.py /my/path/to/data --gpu 0 --model_dir /my/path/to/save/models
In our experiments,
/my/path/to/data contains one
npz file for each subject saved in the variable
We provide a T1 brain atlas used in our papers at
python test_miccai2018.py [gpu-id] [model_dir] [iter-num]
If you simply want to register two images:
register.py. For example, Let's say we have a model trained to register subject (moving) to atlas (fixed). One could run:
python register.py --gpu 0 /path/to/test_vol.nii.gz /path/to/atlas_norm.nii.gz --out_img /path/to/out.nii.gz --model_file ../models/cvpr2018_vm2_cc.h5
For the CC loss function, we found a reg parameter of 1 to work best. For the MSE loss function, we found 0.01 to work best.
For our data, we found
prior_lambda=25 to work best.
In the original MICCAI code, the parameters were applied after the scaling of the velocity field. With the newest code, this has been "fixed", with different default parameters reflecting the change. We recommend running the updated code. However, if you'd like to run the very original MICCAI2018 mode, please use
xy indexing and
use_miccai_int network option, with MICCAI2018 parameters.
The spatial transform code, found at
neuron.layers.SpatialTransform, accepts N-dimensional affine and dense transforms, including linear and nearest neighbor interpolation options. Note that original development of VoxelMorph used
xy indexing, whereas we are now emphasizing
For the MICCAI2018 version, we integrate the velocity field using
neuron.layers.VecInt. By default we integrate using scaling and squaring, which we found efficient.
If you use voxelmorph or some part of the code, please cite (see bibtex):
For the diffeomorphic or probabilistic model:
Unsupervised Learning of Probabilistic Diffeomorphic Registration for Images and Surfaces
Adrian V. Dalca, Guha Balakrishnan, John Guttag, Mert R. Sabuncu
MedIA: Medial Image Analysis. 2019. eprint arXiv:1903.03545
For the original CNN model, MSE, CC, or segmentation-based losses:
VoxelMorph: A Learning Framework for Deformable Medical Image Registration
Guha Balakrishnan, Amy Zhao, Mert R. Sabuncu, John Guttag, Adrian V. Dalca
IEEE TMI: Transactions on Medical Imaging. 2019. eprint arXiv:1809.05231
In our initial papers, we used publically available data, but unfortunately we cannot redistribute it (due to the constraints of those datasets). We do a certain amount of pre-processing for the brain images we work with, to eliminate sources of variation and be able to compare algorithms on a level playing field. In particular, we perform FreeSurfer
recon-all steps up to skull stripping and affine normalization to Talairach space, and crop the images via
((48, 48), (31, 33), (3, 29)).
We encourage users to download and process their own data. See a list of medical imaging datasets here. Note that you likely do not need to perform all of the preprocessing steps, and indeed VoxelMorph has been used in other work with other data.
We present a template consturction method in this preprint:
To experiment with this method, please use
train_img_template.py for unconditional templates and
train_cond_template.py for conditional templates, which use the same conventions as voxelmorph (please note that these files are less polished than the rest of the voxelmorph library).
We've also provided an unconditional atlas in
Explore the atlases interactively here with tipiX!
We recently published a method on deep learning methods for unsupervised segmentation that makes use of voxelmorph infrastructure. See the unified seg README for more information.
2019-11-28: Added a preliminary version of
2019-08-08: Added support for building templates
2019-04-27: Added support for unified segmentation
2019-01-07: Added example register.py file
2018-11-10: Added support for multi-gpu training
2018-10-12: Significant overhaul of code, especially training scripts and new model files.
2018-09-15: Added MICCAI2018 support and py3 transition
2018-05-14: Initial Repository for CVPR version, py2.7
For any problems or questions please open an issue in github.