This is my submission to the MICCAI Educational Challenge 2019.
You can run the notebook on Google Colab or render an already executed version on nbviewer.
NiftyNet is "an open source convolutional neural networks platform for medical image analysis and image-guided therapy" built on top of TensorFlow. Due to its available implementations of successful architectures, patch-based sampling and straightforward configuration, it has become a popular choice to get started with deep learning in medical imaging.
PyTorch is "an open source deep learning platform that provides a seamless path from research prototyping to production deployment". It is low-level enough to offer a lot of control over what is going on under the hood during training, and its dynamic computational graph allows for easy debugging. Being a generic deep learning framework, it is not tailored to the needs of the medical imaging field, although its popularity in this field is increasing rapidly.
One can extend a NiftyNet application, but it is not straightforward without being familiar with the framework and fluent in TensorFlow 1.X. Therefore it can be convenient to implement applications in PyTorch using NiftyNet models and functionalities. In particular, combining both frameworks allows for fast architecture experimentation and transfer learning.
So why not use both? In this tutorial we will port the parameters of a model trained on NiftyNet to a PyTorch model and compare the results of running an inference using both frameworks.