pytorch-mri-segmentation-3D

3D model implementations of:
   DeepLab v3 - paper
   UNET - paper
   HRNet - paper
   EXPNet - experiment models DefaultCNN, PrivCNN

Usage

Data structure

main_folder_path
│
│───scans
│	└───scan1
│		└───pre
│			└───FLAIR.nii.gz
│			wmh.nii.gz
│──────scan2
│		└───pre
│			└───FLAIR.nii.gz
│			wmh.nii.gz
│──────scan3
│		└───pre
│			└───FLAIR.nii.gz
│			wmh.nii.gz

Setup

main_folder_path='../../Data/MS2017test/'
cd utils/

Train landmarks (for histogram normalization)

python trainLandmarks.py --mainFolderPath=$main_folder_path --pcRange=1-99

Resize and normalize scans (histogram normalization & per-subject normalization)

python resizeScans.py --mainFolderPath="$main_folder_path" --postfix=_200x200x100orig --noGIFS --withNorm=1 --size=200x200x100

Generate train/val split

python -c 'import PP;  PP.generateTrainValFile(0.8, main_folder = "'$main_folder_path'", postfix="_200x200x100orig")'

Training

Train architecture 0 (DeepLab v3). Model is saved in train_results/models/

python train.py --useGPU=1 --archId=0 --maxIter=100000 --lr=0.0001 --iterSize=25 --patchSize=60

Train EXPNet model (D1xD2xD3xD4_priv_aspp)

python train.py --experiment=1x1x1x1_1_1 --useGPU=1 --maxIter=100000 --lr=0.0001 --iterSize=25 --patchSize=60

Evaluation

Evaluate model. Results saved in eval_results/

python eval.py --patchPredSize=60 --modelPath=$modelpath --singleEval

Requirements

PyTorch v0.2.0, Python 2.7