Project: Data Augmentation of 3D Medical Images and Landmarks

This project is used for realistic Data Augemntation for brain MRI images using two Libraries: TorchIO and SITK

System requirements

• windows/linux
• Python 3.6 or above
• Pytorch>=1.1
• torchio==0.18.30

Dependencies

• nibabel==3.2.1
• tqdm==4.60.0
• numpy==1.20.1
• torchio==0.18.30
• SimpleITK==2.0.2

Installation

• Install the required libraries using pip install -r requirements.txt

Info

•                              Policy distribution if  ` --policy=all ` is used 
  

There are six types of augmentation available in the project:

Policy	Description
DA-1	One of Spatial Transforms and then one of Non Spatial Transforms
DA-2	One of Non Spatial Transforms
DA-3	One of Spatial Transforms
DA-4	Random Elastic deformation
DA-5	Scaling, Rotation, Translation
ALL	Run all the DA1,DA2,DA3,DA4 with the above given probability distribution

DA-5 only uses the SITK library rest of them uses the TorchIO library Operations available in Spatial/Non Spatial Transforms are :

1. Spatial Transforms : RandomAffine(Translation, Rotataion, Scaling), RandomAnisotropy, RandomElasticDeformation
2. Non Spatial Transforms: RandomGhosting, RandomSpike, RandomBiasField, RandomNoise, RandomMotion, RandomBlur

Note: A particular transformation operation is selected according to the probability assigned in the flow chart

\

•                              Spatial transforms distribution 
  

\

•                              Non Spatial transforms distribution 
  

\

•                              Affine transforms distribution 
  

Run

python TOrchIO+SITK_transforms.py --nifty_input_path <input path of nifty> --fcsv_input_path <input path to the landmarks file in .fcsv> --output_path<output path> --policy <policy no>
eg:
python TOrchIO+SITK_transforms.py --nifty_input_path "../test/testfiles" --fcsv_input_path "../test/testfiles" --output_path "output" --policy "all"

Note: The output Nifty file is saved as <original_filename>DA<policy_no>_<transformation_name>.nii and similarly for the .fcsv file for the given eg. case, Policy 4 was selected and Random Elastic Transformation was applied, so the file was saved as ref_1_OAS30591_MR_d0914_v4_DA4_RandomElasticDeformation.nii

Examples

1. Performs DA1-4 with the associated policies -----> set --policy=all
2. Perform a given Data augmentation policy -----> set --policy=<policy no>
3. To perform data augemntation using the SITK module(for rotation,translation,scaling) -----> set --policy=5

Runtime

Pls note the Runtime for the given system Configuration

Hardware/Software	Specs
RAM	16GB
CPU	Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz (8 CPUs), ~2.8GHz
Platform	Windows 10

Transformation type	Time/image
Spatial Transdforms	~ 5 mins
Non Spatial Transforms	~ 3 mins
Affine Transforms(using SITK)	~2 mins

Create your own policies

To create your own policies read this TorchIO documentation and modify the Policy() class in TOrchIO+SITK_transforms.py