lab2im

We provide here a generative model that synthesises images by sampling a Gaussian Mixture Model (GMM) conditioned on label maps. For each image, the GMM parameters are sampled from prior distributions of predefined hyperparameters, thus leading to differences in contrast for all synthesied images.

Aditionally to sampling the GMM, data augmentation is performed during the generation process. This includes spatial deformation of the label maps, random cropping, random blurring, and intensity augmentation steps such as bias field corruption, and gamma augmentation.

Importantly, we provide all the steps of the generative model as separate functions, so that user can easily build their own model. We encourage you to have a look at the SynthSeg package, which uses the provided functions to build a more complete generative model!

The model is implemented in keras with a tensorflow backend, and relies on the neuron package [1,2].

The following figure shows synthetic brain MRI scans generated whith lab2im. It also illustrates the effect of using prior distributions when sampling the GMM parameters.

---

Try it in two lines !

from lab2im.image_generator import ImageGenerator

brain_generator = ImageGenerator("./tutorials/data_example/brain_label_map.nii.gz")
im, lab = brain_generator.generate_image()

---

Content

• lab2im: this is the main folder containing the lab2im model and separate functions to manipulate tensors:

  • lab2im_model.py: contains the generative model lab2im_model.

  • image_generator.py: contains the class ImageGenerator a wrapper around lab2im_model. One can simply generate new images by instantiating an object of this class, and call the method generate_image().

  • sample_gmm.py: contains the function sample_gmm_conditioned_on_labels to sample a GMM of defined parameters conditionally on a label map.

  • spatial_augmentations.py: contains functions to spatially augment the input label maps, such as deform_tensor for linear and elastic deformation, random_cropping, or label_map_random_flipping.

  • intensity_augmentation.py: functions to perform intensity data augmentation such as bias_field_augmentation, gamma_augmentation, and min_max_normalisation.

  • edit_tensors.py: contains several types of functions to edit keras/tensorflow tensors. It notably has the blurring function blur_tensor along with get_gaussian_1d_kernels, which enables to create blurring kernels as tensors. It also has a function resample_tensor to resample a tensor to the desired resolution.

  • edit_volumes.py: contains numerous functions to preprocess volumes and label maps (given as numpy array, not tensors!), such as masking, cropping, resampling, smoothing,... These functions are classified by volume type (volumes/label maps), and by input type (numpy arrays/ paths to volumes).

  • utils.py: contains all the utility functions used in this repository.

• tutorials: simple scripts explaining how to easily generate images. Start with simple_example.py, which shows how to generate images in 2 lines. Then have a look at random_contast_generation.py that introduces some functionalities of ImageGenerator. Finally, t1w_generation.py examplifies how to impose prior distributions to the Gaussian distributions of the GMM, in order to sample image of desired intensity distributions.

• ext: contains modules, especially the modified version of neuron

---

Requirements

All the requirements are listed in requirements.txt. We list here the important dependencies:

• tensorflow-gpu 2.0
• tensorflow_probability 0.8
• keras > 2.0
• cuda 10.0 (required by tensorflow)
• cudnn 7.0
• nibabel
• numpy, scipy, sklearn, tqdm, pillow, matplotlib, ipython, ...

Because some changes were made compared to the current version of neuron, the package is already included in this repository.

---

Citation/Contact

This code is under Apache 2.0 licensing.
If you use it, please cite the following paper:

A Learning Strategy for Contrast-agnostic MRI Segmentation
Benjamin Billot, Douglas N. Greve, Koen Van Leemput, Bruce Fischl, Juan Eugenio Iglesias*, Adrian V. Dalca*
*contributed equally
MIDL 2020
[link | arxiv | bibtex]

If you have any question regarding the usage of this code, or any suggestions to improve it you can contact us at:
benjamin.billot.18@ucl.ac.uk

---

References

[1] Anatomical Priors in Convolutional Networks for Unsupervised Biomedical Segmentation
Adrian V. Dalca, John Guttag, Mert R. Sabuncu
CVPR 2018

[2] Unsupervised Data Imputation via Variational Inference of Deep Subspaces
Adrian V. Dalca, John Guttag, Mert R. Sabuncu
Arxiv preprint 2019