BraTS2018_NvNet

BraTS_NvNet

This is an unofficial Pytorch implementation of BraTS_NvNet, based on Andriy Myronenko’s paper on MICCAI_BraTS_2018_proceedings_shortPapers

Overview

BraTS2018: Multimodal Brain Tumor Segmentation Challenge 2018

The participants are called to address this task by using the provided clinically-acquired training data to develop their method and produce segmentation labels of the different glioma sub-regions. The sub-regions considered for evaluation are: 1) the “enhancing tumor” (ET), 2) the “tumor core” (TC), and 3) the “whole tumor” (WT).

Segmentation Labels:

• WT: 2
• TC: 1
• ET: 4

Network

A combination of Vnet and VAE(variation auto-encoder).

Table 1. Encoder structure, where GN stands for group normalization(with group size of 8), Conv - 3*3*3 convolution, AddId - addition of identity / skip connection. Repeat column shows the number of repetitions of the block.

Table 2. Decoder structure, where Conv1 stands for 1*1*1 convolution, UpLinear - 3D linear spatial upsampling.

Table 3. VAE decoder branch structure, where Dense stands for fully connected layer.

sample ~

Loss

1. The loss function consists of 3 terms:

   

2. is applied to the decoder output to match the segmentation mask :

   

   where summation is voxel-wise, and the is a small constant to avoid zero division.

3. is appiled on the VAE branch output to match the input image :

   

4. is standard VAE penalty term, a KL divergence between the estimated normal distribution and a prior distribution , which has a closed form representation:

   

   where N is total number of image voxels.

5. The hyper-parameter weight of 0.1 was empirically set to provide a good balance between dice and VAE loss terms.

Dependencies

python: 3.6.2
pytorch: 0.4.1
pytables: 3.4.4
numpy: 1.13.1
tensorboardX: 1.4
nibabel: 2.3.0
nilearn: 0.4.2
tqdm: 4.26.0
pickle: 0.7.4

Usage

1. data preprocessing:

• change the directory of brats 2018 training dataset and validation dataset
• set dataset format
• run the script：

  python3 data_preprocess.py

2. train model:

• set training data file path
• set training parameters
• run the script:

  python3 main.py

3. predict:

• set validation data file path
• set model file path
• run the script:

  python3 predict.py

4. make submission:

• set prediction file path
• set reconstruction parameters
• run the script:

  python3 make_submission.py

Implementation Details

• NvNet
• Loss
• DataLoader
• Train
• Predict
• Data Preprocess
• Make Submission