jojojoshuali / 3d_cnn_brain_lesion_segmentation

Detecting Brain Lesions using 3D CNN Techniques

UCLA_CS168_S18

Detecting Brain Lesions using 3D CNN Techniques

Click here to view our poster...

https://github.com/LawrenceXu13467/3D_CNN_Brain_Lesion_Segmentation https://lawrencexu13467.github.io/3D_CNN_Brain_Lesion_Segmentation/

All the code for training code and contructing the model is written in python.

required python library:

Theano

NiBabel

Parallel Python

scipy

numpy

six

nibabel

keras

pytables

nilearn

SimpleITK

nipype

use following command to install the python library

$ sudo pip3 install NAMEOFTHELIBRARY

CUDA is required to run the code. Go on the Cuda website: https://developer.nvidia.com/cuda-90-download-archive to install the CUDA9.0 for your computer

please make sure the CUDA is installed correctly

nvcc --version

First of all, the image from the dataset is required to be preprocessed to fit the both of the 3D CNN models.

In regard to 3D unet, the main issue is to correct the bias before the training to prevent the supervising algorithm in the model from generalizing beyond the training set by using ANTs N4BiasFieldCorrection.

For the DeepMedic model, the default configuration on scans is 200x200x200 and in order to fit the model, the format of the images should be NIFTI. Also, it is required that the number of voxel per dimension should be the same for each image. Due to DeepMedic’s characteristic, the background of each images should be set to zero for training.

The first 3D CNN model we choose is referencing from the 3D unet. The model is build from the keras library from python, which provides many useful class to construct the 3D unet model. The model is first applied with two types of levels of convolution blocks, the max pooling and up-convolution which both are the classes provided the keras library. The following is the library we use to construct for 3DUnet

from keras.layers import Conv3D, MaxPooling3D, UpSampling3D, Activation, BatchNormalization, PReLU, Deconvolution3D

Each levels is based on the depth and number of the base filter. Even though a deeper depth could ensure the model to be precise, it is important to choose a appropriate depth size and number of base filter for our model since the more of the depth or number of base filter, the more memory that the training will take since the 3D unet is appending the multiple filters for each layer based on the depth parameter and initial base of the filters, i.e. a inappropriate depth or number of base filter could lead exceeding memory allocation issues. For us, the parameter is set to be 32-base-filters and the depth to be 4. Finally, by applying Conv3D from Keras, we build the final convolutional neural network for our dataset.

The second 3D CNN model we choose is referencing from the 3D CRF model with application of the residue connection, specifically the DeepMedic model. The DeepMedic model is mainly constructed through the Theano library provided from the Python. According to the paper, DeepMedic is a double pathway architecture for multi-scale processing. To implement, the feature map per layer is set to be [30,40,40,50] and dimension of kernel per layer is set to be [[5,5,5],[5,5,5],[5,5,5]]. It is important to set the appropriate batch size while we create our model. Even though the large batch size allows us to process the training simultaneously on GPU which allows faster processing, such process would introduce a large amount of computation and requires extra amount of memory location which could potentially introduce crush of the program. After setting up the model’s configuration, we then begin to build our actual model by make the normal pathway of the CNN and then apply subsampled pathway of CNN which creates essentially a upsampling layer. Then the algorithm will be concatenated

reference source code to our research:

DeepMedic source code: https://github.com/Kamnitsask/deepmedic

3DUnet source code: https://github.com/ellisdg/3DUnetCNN

where we made modification on these source code to fit in our own data.