Deep Learning Flows for Neuroimaging

This document describes the use of the dl4ni package, Deep Learning Flows for Neuroimaging.

Introduction

Recent years have seen the rise of Deep Learning applied to Neuroimgining problems. Some concepts and state-of-the-art are reviewed here and here.

There are a lot of packages in the R ecosystem focused on several aspects of Neuroimage processing, many of them listed in the CRAN Task View: Medical Imaging section: oro.nifti, oro.dicom, RNifti, fmri, divest (among others) for reading/writing of images; ANTsR, fslr for more advanced processing methods; RNiftyReg specialized in registration, mritc focused in image segmentation. Many more, for DTI, fMRI, molecular imaging, and visualization are available in CRAN.

For Deep Learning, there are basically 4 packages maintained by diverse organizations: mxnet from Apache MXNet, h2o from H2O.ai and tensorflow and keras from Google and RStudio. All of these packages are available at CRAN.

The lack of packages to apply Deep Learning techniques to Neuroimaging pipelines inspired the creation of the dl4ni package.

Main Features

This package has been written and is developed with some key features in mind:

• Simple definition of Neuroimaging flows, which can incorporate both R functions (including those from other packages, such as ANTsR or fslr) and Deep Learning models.
• Simple definition of Deep Learning models. Many wrappers to common models (U-Net, SegNet ...) have been included to ease the definition of models.
• Highly customizable models, allowing concatenation and merging of models, layers, branching...
• Creation of a common interface for all models built with this package, regarding training, testing and inference. This is achieved by the use of R6 classes.
• Tackle different types of problems: classification, labelling or regression.
• Native 3D, although vector representations of images are allowed (and, in some cases, adviced).
• Simple memory management. Users can impose memory constraints when training a model, batch size is automatically adjusted.
• Dataset repositories. Some datasets regarding typical Neuroimaging problems (brain extraction, segmentation, parcellation...) are released in public repositories, and a mechanism to get these datasets into our local system to be able to train models is also included in the package.

Requirements

This package is based in the keras package, so a working installation of both tensorflow and keras is a minimum requirement. Check the package website for detailed instructions on how to get those packages installed and running.

Installation

Currently, this package is not in CRAN. The only way to get this package installed is by obtaining the source and building and installing the package in a local system.

#!bash
R CMD INSTALL /path/to/dl4ni

Examples

Feel free to use any of the provided demos in the package. Just load the package and list the available demos as:

#!R
library(dl4ni)

list.files(system.file("demo", package = "dl4ni"))

And open the one you like most!

Additional Information

You can find more detailed explanations on flows here and on models here. Also, package vignettes are available.