#Random Forest for Multiple Sclerosis Lesion Segmentation - MS-challenge-2016
Version 1.0.0
Author: Francisco Javier Vera Olmos mail:javier.vera@urjc.es
This project is based on Python 2.7 and Matlab Compiler Runtime to use SPM toolbox.
This paper explain how this software works: https://www.overleaf.com/read/jjkhxzszpdhv
To use this software you need the followings python packages:
MedPy
NumPy
scikit-learn
SciPy
matplotlib
XlsxWriter(optional)
The easiest way to install them is using Conda, wich install most of the package, and then use pip to install MedPy, scikit-learn and XlsxWriter(optional).
Also it is mandatory needed to have installed the Matlab Compiler Runtime(MCR) and a SPM standalone installation compatible with that MCR.
To execute the program you need to parse the path of your SPM standalone isntallation and (only if you use a Linux or a MacOS systems) you have to parse the path to your MCR.
Also you need the classifiers ,which are larges files, they are avalibles in this link: classifiers
Once downloaded, move it to the root of the project and uncompress it(take in count it will take 10 Gb of space more or less). You should have in the root of the project a folder named classifiers.
The inputs are the followings:
positional arguments: t1_raw---> path of the T1-w volume
flair_raw ---> path of the FLAIR volume
t1_pp ---> path of the T1-w volume
flair_pp ---> path of the FLAIR volume
brain_mask BRAIN_MASK ---> path of the brain-mask volume
spm_path ---> path to spm.sh if linux or spm.exe if windows
output_folder ---> path where outputs and intermediate files will be stored
optional arguments:
-h, --help ---> show this help message and exit
-mcr_path MCR_PATH ---> matlab compiler runtime path
-t2_raw T2_RAW ---> path of the T2-w volume
-dp_raw DP_RAW ---> path of the DP volume
-gado_raw GADO_RAW ---> path of the T1-Gado volume
-t2_pp T2_PP ---> path of the T2-w volume
-dp_pp DP_PP ---> path of the DP volume
-gado_pp GADO_PP ---> path of the T1-Gado volume
For now two modes are implemented :
1- Using T1 and FLAIR sequence
2- Using T1, T2, FLAIR, Proton Density (DP) and Gadoline T1 (GADO)
It is mandatory to provide the paths of the raw images and the preprocessed volumes(skull stripped and bias corrected) of every sequence.
The outputs generated are stored in two folders: "intermediates" and "results".
In "intermediates" you have all the files and vols that must be generating during the running of the software. This intermediates files depends on the inputs provided and they can be the followings:
Folder descriptors --> It contains the descriptor generated to use the Random Forest.
Folder results_csf_ext --> It contains the segmentation of the CSF exterior.
Folder results_rf --> It contains the probabilistic mask generated by the Random Forest.
File brain_mask.nii.gz --> Brain mask it is generated only if not provided.
File c1[nameofyourT1].nii --> CSF segmentation.
File c2[nameofyourT1].nii --> WM segmentation.
File c3[nameofyourT1].nii --> GM segmentation.
File [nameofyourT1]_corrected.nii.gz --> T1 with intensity corrected.
File [nameofyourT2]_corrected.nii.gz --> T2 with intensity corrected.
File [nameofyourFLAIR]_corrected.nii.gz --> FLAIR with intensity corrected.
File [nameofyourDP]_corrected.nii.gz --> DP with intensity corrected.
File [nameofyourGADO]_corrected.nii.gz --> GADO with intensity corrected.
File [nameofyourT1]_seg8.mat --> mat generated by SPM.
File [nameofyourT1].nii --> uncompressed T1 needed for SPM,it is generated if data is given compressed.
File gen_batch.m --> batch generated to run in SPM.
In "results" folder you will have the final lesion mask generated.
Example of usage: python ms_run.py 3DT1.nii.gz 3DFLAIR.nii.gz T1_preprocessed.nii.gz FLAIR_preprocessed.nii.gz spm12_r6685\spm12\spm12_win64.exe output_path
#Summary
Multiple Sclerosis (MS) is a chronic, inflammatory and demyelinating disease that primarily affects the white matter of the central nervous system. Automatic segmentation of MS lesions in brain MRI has been widely investigated in recent years with the goal of helping MS diagnosis and patient follow-up. It offers an attractive alternative to manual segmentation, which remains a time-consuming task and suffers from intra- and inter-expert variability. We propose a new approach that uses a Random Forest (RF) classifier. Its input has been filtered with a threshold based on the gray matter (GM) distribution and uses several features that take into account voxel and context information. A Markov Random Field (MRF) post processing algorithm has been applied to make lesions grow through probable neighborhoods. Our approach will be used in the MS challenge 2016, that will take place in October during MICCAI 2016. Therefore, in order to train and test the method we use the database provided from the challenge, which consists in 15 subject from 4 different centers, imaged on 1.5T or 3T scanners.The provided MR sequences include: 3D FLAIR, 3D T1-w, 3D T1-w GADO, 2D DP and 2D T2. In this data set each patient has being manually annotated by seven experts. We found that segmentation results are maximized by using all available sequences, but the FLAIR volume provides most of the information. Thus our method can work only with the T1-w volume to segment the tissues and the FLAIR volume to extract the most important features, still, its performance improves slightly if more types of sequences are available.
#FAQ
In the future I will add support for just raw images and implement the skull stripping and the bias correction, but FSL which is the most used in skull stripping is only native for linux and I want to keep this project in mostly python code and for all plataforms, so it's going to take time.