Nobrainer-zoo is a toolbox with a collection of deep learning neuroimaging models that eases the use of pretrained models for various applications. Nobrainer-zoo provides the required environment with all the dependencies for training/inference of models. The only software needed is singularity or Docker.

To use the Nobrainer-zoo,

git clone https://github.com/neuronets/zoo.git
cd zoo
pip install -e .

Models should be refrenced based on their organization, and model name (neuronets/brainy). The trained models are version controled and one model might have different version. Therefore for inference, the model version also needs to be specified(neuronets/brainy/0.1.0). Some models (kwyk and braingen) also have various types which means there was different structural chracteristic during training that leads to different trained models. Run help to see the functions and each function's options.

nobrainer-zoo --help
nobrainer-zoo predict --help
nobrainer-zoo train --help

• brainy: 3D U-Net brain extraction model (available for training and inference)
• ams: 3D U-Net meningioma segmentation model (available for training and inference)
• SynthSeg: Contrast and resolution invariant 3D brain segmentation model (available for inference)
• SynthSR: Contrast, resolution and orientation invariant MRI/CT hyper resolution model (available for inference)

List of models which will be added in near future can be find here. You can suggest a model here.

Note: models are distributed under their original license.

Inference with default options,

nobrainer-zoo predict -m neuronets/brainy/0.1.0 <path_to_input> <path_to_save_output>

nobrainer-zoo predict -m UCL/SynthSeg/0.1 <path_to_input> <path_to_save_output>

pass the model specific options with --options argument to the model.

nobrainer-zoo predict -m neuronets/brainy/0.1.0 <path_to_input> <path_to_save_output> --options verbose block_shape=[128,128,128]

nobrainer-zoo predict -m UCL/SynthSeg/0.1 <path_to_input> <path_to_save_output> --options post=<path_to_posteriors>

For training with sample dataset you do not need to pass any dataset pattern.

nobrainer-zoo train -m neuronets/brainy

To train the network with your own data pass the dataset pattern in the form of tfrecords.

nobrainer-zoo train -m neuronets/brainy "<data_train_pattern>" "<data_evaluate_pattern>"

Other parameters can be changed by providing a spec file or changing them with cli command.

nobrainer-zoo train -m neuronets/brainy --spec_file <path_to_spec_file>

nobrainer-zoo train -m neuronets/brainy --train epoch=2