Prediction of Modified Rankin Scale (mRS) in Stroke Patients with Tractographic Features

This repository extracts the tractograpic feature, other first-order features and the state-of-the-art feature from the stroke lesion. A random forest regressor is used with one type of feature to predict the modified Rankin Scale of stroke patients.

Citations

Predicting the overall survival of brain tumor patients using tractographic feature

Kao, Po-Yu, et al. "Brain tumor segmentation and tractographic feature extraction from structural mr images for overall survival prediction." International MICCAI Brainlesion Workshop. Springer, Cham, 2018.

Predicting the clinical outcome of stroke patients using tractographic feature

Kao, Po-Yu, et al. "Predicting Clinical Outcome of Stroke Patients with Tractographic Feature." International MICCAI Brainlesion Workshop. Springer, Cham, 2019.

Dataset

Ischemic Stroke Lesion Segmentation (ISLES) 2017

Dependencies

Python 3.6

Required Python Libraries

SimpleITK, scipy, skimage

Required Softwares

For image registration, you need to download FSL.

For fiber tracking and building connectivity matrix, you need to download DSI Studio.

How to run the codes

1. Change the pathes inside paths.py

set isles2017_dir to the path you store the clinical parameters file (ISLES2017_Training.csv)

set isles2017_training_dir to the path you save ISLES2017 training data (ISLES2017/train)

set mni152_1mm_path to the path store the MNI152_T1_1mm_brain.nii.gz

set dsi_studio_path to the dsistudio directory

2. Run registerBrain.py

This script registers the MR-ADC image and the brain lesion from the subject space to MNI152 1mm space.

The outputs:

ADC_MNI152_1mm.nii.gz, ADC_MNI152_1mm_invol2refvol.mat, and ADC_MNI152_1mm_refvol2invol.mat under ADC's directory

OT_MNI152_1mm.nii.gz and OT_prob_MNI152_T1_1mm.nii.gz under brain lesion's directory

3. Run fiber_tracking.py

This script generates the fiber tracts for the subject.

We seed in the whole brain region and find the fiber tracts passing through the lesion region

The outputs:

end-type connectivity matrix and pass-type connectivity matrix

end-type connectogram and pass-type connectogram

end-type network measures and pass-type network measures

4. predict.py

Perform mRS prediction on features extracted from the lesion region with leave-one-out cross-validation on the ISLES2017 training dataset

5. extract_oskar_features.py (run in python 2.7)

Required python libraries:

nibabel, medpy, skimage

Extract the features descirbe in the ISLES2016 winning paper

This script extracts 1662 features for each subject.

utils.py

Provide you different types of features and tools for processing the brain images

analysis.py

Provide you the confusion matrix and the p-value

heatmap.py

Create the heatmap of the stroke lesion in MNI space