Data augmentation

[ravih18]

One of the current limitation of ClinicaDL is the miss of data augmentation possibilities. We know that in deep learning applied to medical imaging, data augmentation is almost a standard procedure in images pretreatment.

Prerequisite

The main issues are:

• How to add it? Which library do we use?
• Will it be performed during training (during data loading)? Or in the extraction step ?
• Or do we add a new command / new step for those who want to perform data augmentation ?

An other point is to define which transformations we would like to add? This will require to identify all the basic and most used transformations in the literature. Then we can imagine that users will have to add their own transformations if needed.

Tools

The torch.io seems to be the most adequate tool for data augmentation. It is already part of the official PyTorch Ecosystem and includes all the functionalities that we would like to integrate in ClinicaDL.

Moreover, it could also be used to improve the data loading, preprocessing and extraction (for patch) that are already available in ClinicaDL.

Existing project

Monai is another deep learning for medical imaging and it already integrates data augmentation. It will be interesting to see how it is done.