IDH_Prediction

Use of Deep Learning to Predict IDH status of gliomas from MR Imaging

1. Chang K, Bai HX, Zhou H, Su C, Bi WL, Agbodza E, et al. Residual Convolutional Neural Network for Determination of IDH Status in Low- and High-grade Gliomas from MR Imaging. Clin Cancer Res [Internet]. American Association for Cancer Research; 2017 [cited 2017 Nov 29];clincanres.2236.2017. Available from: http://www.ncbi.nlm.nih.gov/pubmed/29167275

This is a repository for pre-processing of MR images of gliomas. The input data is assumed to be 4 modality MR imaging (T2 FLAIR, T2, T1 pre-contrast, T1 post-contrast) in dicom format.

The order in which the scripts should be run are

1. ResampleRegister.m - Registration and isotropic resampling
2. n4_skullstrip.py - n4 bias correction and skullstripping
3. normalize_intensity.py - Normalize image intensity by subtracting median and dividing by interquartile range of normal brain
4. compile_patientsamples.py - Extract patient image samples, age, and idh status and compile them into numpy files.
5. predict.py - Predict IDH status using modality network models, combining outputs with age in a logistic regression

Step 1) is a MATLAB script while the others are python scripts. Step 5) is written in Keras 2 with TensorFlow backend. The trained models can be downloaded here: https://www.dropbox.com/sh/enfdwh8qh8x5yro/AADtOMbUqfmUtEGA9SDBwGeja?dl=0

dependancy versions: keras 2.1.3 tensorflow 1.2.0 scikit-learn 0.14.1 joblib 0.7.1