Neurite

A neural networks toolbox with a focus on medical image analysis in tensorflow/keras for now.

⚠️ Warning: neurite is under active development. We are in the process of finalizing the structure for PyTorch. Please keep this in mind when using the library, as interfaces may change.

Install

To use the Neurite library, either clone this repository and install the requirements listed in setup.py or install directly with pip.

pip install neurite

Main tools

• layers: various network layers, sparse operations (e.g. SpatiallySparse_Dense), and LocallyConnected3D currently not included in keras
• utils: various utilities, including interpn: N-D gridded interpolation, and several nonlinearities
  • model: stack_models: keras model stacking
  • vae: tools for analyzing (V)AE style models
  • seg: segmentation tools
• models: flexible models (many parameters to play with) particularly useful in medical image analysis, such as UNet/hourglass model, convolutional encoders and decoders
• generators: generators for medical image volumes and various combinations of volumes, segmentation, categorical and other output
• callbacks: a set of callbacks for keras training to help with understanding your fit, such as Dice measurements and volume-segmentation overlaps
• dataproc: a set of tools for processing medical imaging data for preparation for training/testing
• metrics: metrics (most of which can be used as loss functions), such as Dice or weighted categorical crossentropy
• plot: plotting tools, mostly for debugging models

Papers:

If you use this code, please cite:

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

If you are using any of the sparse/imputation functions, please cite:

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

Development:

We welcome contributions; please make sure your code respects pep8, except for E731,W291,W503,W504, by running:
pycodestyle --ignore E731,W291,W503,W504 --max-line-length 100 /path/to/neurite
Please open an issue [preferred] or contact Adrian Dalca at adalca@csail.mit.edu for question related to neurite.

Use/demos:

Parts of neurite were used in VoxelMorph and brainstorm, which we encourage you to check out!