devhliu / mip-deepstroke

Deep learning network for predicting lesions in stroke

MIP-DeepStroke

Deep learning network for predicting lesions in stroke.

This project aims to predict stroke lesions from T2-weighted, CT-Perfusion and Diffusion-weighted images. This research was supported by the Medical Image Processing Lab (MIPLAB) in Campus Biotech Geneva.

Requirements

• Tensorflow(-GPU)
• Keras

How to I use this code?

This code is relatively generic and might be used for any binary segmentation task.

Dataset creation

First, the different input and output images should be organised in the following manner. The images should stored by patient in .nii format.

~/training_data/patient_id/[input_modalities1.nii, input_modalities2.nii, output_mask.nii]

Now that the folder is well organised, you can run the following script to create a dataset of patches from the images.

• For 2D patches: python ~/MIP-Deepstroke/Unet2D/create_dataset.py -d ~/training_data -p 512x512 for 512x512 patches.
• For 3D patches: python ~/MIP-Deepstroke/UnetCT/create_dataset.py -d ~/training_data -p 32x32x32 for 32x32x32 (3D) patches.

By default, the data is saved under ~/Data/{ddmmyy-hhmm}/{patch_size}.

Training

To train a model on the data just created, use the following command:

• For 2D : python ~/MIP-Deepstroke/Unet2D/train_2D.py -d {DATAPATH} -i {INPUT_MODALITY1} -i {INPUT_MODALITY2} -o {OUTPUT_MODALITY}
• For 3D : python ~/MIP-Deepstroke/Unet2D/train.py -d {DATAPATH} -i {INPUT_MODALITY1} -i {INPUT_MODALITY2} -o {OUTPUT_MODALITY}

Example : python ~/MIP-Deepstroke/Unet2D/train.py -d ~/Data/010119-1000/32x32x32 -i T2 -i TRACE -o LESION

By default, models are saved under ~/Models/{date}/checkpoints.

Note

This README is still under construction.

External code

• 2D-Unet : https://github.com/nabsabraham/focal-tversky-unet
• 3D-Unet : https://github.com/ellisdg/3DUnetCNN