3D data augmentation

[reubano]

Overview

The majority of the solutions to #1, #2, #3 have similar augmentation pipelines (e.g. grt123, Therapixel, Daniel Hammack, etc.). These pipelines mainly consists of Hounsfield scaling, lung segmentation (#120), CT re-orientation, and other resource consumptive spacial operations such as zoom, rotate, shear, shift, flip, etc.

While the Keras DataGenerator can yield a processed patches via affine transformations, it is inefficient and doesn't work with 3D data.

Expected Behavior

A DataGenerator should enable people to perform data augmentation operations such as zoom, rotate, shear, shift, flip, etc on CT scans

Technical details

• extend the DataGenerator to deal with 3D volumetric data such as CT scans

Acceptance criteria

• efficiently zoom, rotate, shear, shift, and flip CT scans

NOTE: All PRs must follow the standard PR checklist.