MRI脑部提取-基于3D卷积神经网络实现的MRI脑部提取算法-附项目源码-优质项目实战.zip

Installation We strongly suggest using a GPU if speed is of necessity, as speedups of ~ 40x and more over CPU mode are typical. Nevertheless, the code will run on both CPU or GPU without modification. If you chose to use a GPU, make sure to use a NVIDIA model (these are the only cards supporting CUDA). #### Prerequisites ** Python 2: We recommend the use of Anaconda (https://www.anaconda.com/products/individual#Downloads), especially for Windows users. ** Theano -- This is a GPU-toolkit which our code uses to build and train convolutional neural networks. It is straight forward to get working with GPUs on Linux, but slightly harder on Windows E.g. see: https://pypi.org/project/Theano/ A quick summary to install Theano for Windows users: 1) Install Anaconda (Python 2.7+ but not 3.x, x64) 2) [in console] conda update conda 3) [in console] conda update --all 4) [in console] pip install Theano 5) [in console] conda install mingw libpython for GPU on Windows: 1) install MS Visual Studio (sadly the express version currently does not work) 2) install CUDA 3) add to Path: C:\Program Files (x86)\Microsoft Visual Studio 11.0\VC\bin [or equivalent] C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin [or equivalent] C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\libnvvp [or equivalent] It is still tricky to get it to work, as Theano doesn't properly support Windows ** Linux/Windows: [optional] install NVIDIA cuDNN -- better convolution implementation for GPU users, large speed-gain for training and testing (dowload is free but requires registration) ** nibabel -- allows python code to load nifti files: [in console] pip install nibabel ** h5py -- file format 1) [in console] conda install h5py or pip install h5py #### Configuring Theano In order to configue theano to automatically use the GPU, create a file named .theanorc (.theanorc.txt for Windows users) in your user's home directory and add the following three lines to it -- this is a minimalistic setup, many more options are available: [global] floatX=float32 device=gpu0 You can replace "gpu0" with e.g. "cpu", or select another gpu using "gpu1" (provided you have two GPUs), etc. #### Where can I obtain the data https://www.nitrc.org/projects/ibsr http://www.oasis-brains.org/app/action/BundleAction/bundle/OAS1_CROSS http://loni.usc.edu/atlases/Atlas_Detail.php?atlas_id=12 If you use this data please cite the corresponding publications and comply with the indicated copyright regulations stated. Please understand, that the brain tumor data set used in our publication cannot be made publicly available. ## III. Data preprocessing Data pre-processing is not required, as long as the data does not contain artifacts/NaN entries etc. but it will most likely improve the results. The provided code automatically standardizes the data (zero mean, unit variance per volume), thus only nonlinear data pre-processing operations have any effect. If you use data from different scanners that produce data with varying orientation it might be necessary to transform all your data to a common orientation. For instance, this might be achieved using fslreorient2std [http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSL]. ## IV. Examples #### Brain mask prediction using an already trained CNN ``` python deep3Dpredict.py --help usage: deep3Dpredict.py [-h] -i I [I ...] [-o O] [-n N] [-c C] [-f F] [-prob PROB] [-gridsize GRIDSIZE] Main module to apply an already trained 3D-CNN to segment data optional arguments: -h, --help show this help message and exit -i I [I ...] Any number and combination of paths to files or folders that will be used as input-data for training the CNN -o O output path for the predicted brain masks -n N name of the trained/saved CNN model (can be either a folder or .save file) -c C Filter connected components: removes all connected components but the largest two (i.e. background and brain) [default=True] -f F File saving format for predictions. Options are "h5", "nifti", "numpy" [default=nifti] -prob PROB save probability map as well -gridsize GRIDSIZE size of CNN output grid (optimal: largest possible divisor of the data-volume axes that still fits into GPU memory). This setting heavily affects prediction times: larger values are better. Values that are too large will cause a failure due to too little GPU-memory. python deep3Dpredict.py -n OASIS_ISBR_LPBA40__trained_CNN.save -i /home/share/brain_mask/__NEW__/ibsr_data/02/IBSR_02_ana.nii.gz -gridsize 16 ``` #### Train a new CNN (with your data) ``` python deep3Dtrain.py --help usage: deep3Dtrain.py [-h] -data DATA [DATA ...] -labels LABELS [LABELS ...] [-lr LR] [-name NAME] [-convert_labels CONVERTLABELS] Main module to train a 3D-CNN for segmentation optional arguments: -h, --help show this help message and exit -data DATA [DATA ...] Any number and combination of paths to files or folders that will be used as input-data for training the CNN -labels LABELS [LABELS ...] Any number and combination of paths to files or folders that will be used as target for training the CNN (values must be 0/1) -lr LR initial learning rate (step size) for training the CNN (default: 10^(-5)) -name NAME name of the model (affects filenames) -- specify the same name when using deep3Dtest.py -convert_labels CONVERTLABELS if labels are not binary: this will convert values >1 to 1 -data_clip_range [LOWER UPPER] [optional] specify two values For each data file you have to supply an associated label file. The file names should indicate their relationship such that alphabetical ordering results in a correct matching of the corresponding data and label files, e.g. [vol1_data.nii.gz, vol2_data.nii.gz, ...] <-> [vol1_label.nii.gz, vol2_label.nii.gz, ...] python deep3Dtrain.py -data data/ -labels labels/ ```