silvanac/uniandes-gpu repository overview

This is the implementation -at test time- in a Docker container (with GPU) of the algorithm described in "Brain Tumor Segmentation and Parsing on MRIs using Multiresolution Neural Networks" use to participate in the Multimodal Brain Tumor Segmentation Challenge 2017. This algorithm is able to generate a tumor segmentation on any multimodal brain scan that is preprocessed like a BraTS test subject. The 3 models used were trained in the BraTS2017 training dataset.

Method: We introduce a neural network for brain tumor semantic segmentation that parses their internal structures and is capable of processing volumetric data from multiple MRI modalities simultaneously. As a result, the method can learn from small training datasets. We develop an architecture that has four parallel pathways with residual connections. It receives patches from images with different spatial resolutions and analyzes them independently. The results are then combined using fully-connected layers to obtain a semantic segmentation of the brain tumor. We evaluated our method, using the average of 3 of our models (trained with 35, 38 and 47 epochs), in the 2017 BraTS Challenge dataset, reaching average dice coefficients -in the Whole Tumor task- of 89%, 88% and 86% over the training, validation and test images, respectively.

Data access: Our method reads available volumes as they are organized in Brats2017 dataset, in files per patients with the following characteristics: NIfTI GZIP Compressed Tar Archive (.nii.gz) format, with all header information except the spatial resolution removed, and the individual volumes will be named ‘fla.nii.gz’, ‘t1c.nii.gz’, ‘t1.nii.gz’, ‘t2.nii.gz’. They should be co-registered, skull-stripped, resampled to 1mm^3 isotropic resolution, and aligned to the SRI space. Resulting files: The input data files (‘fla.nii.gz’, ‘t1c.nii.gz’, ‘t1.nii.gz’, ‘t2.nii.gz’) volumes, will be linked to /data and the resulting files will be placed in /data/results. Our method will produce 5 files per patient:

• Segmentation: e.g. "tumor_uniandes_class.nii.gz"
• Probability maps per label class (0,1,2,4): e.g. "tumor_uniandes_prob_1.nii.gz"

Docker info: The image is available in Docker Hub, the size of the image is approx. 3.45Gb, use the following command to pull it:

              nvidia-docker pull silvanac/uniandes-gpu

Running the Docker: We provide a Docker run command with GPU usage.

  NV_GPU=1 nvidia-docker run -v <directory>:/data -it silvanac/uniandes-gpu /python main.py

Using this script, we did a running example with 3 patients from the 2017BraTS validation dataset in a virtual machine with 40 CPUs and 126Gb of memory. The result was that the process tended to use all available CPUs, fluctuated occupying between 7Gb and 20Gb of memory and lasted between 30 and 55 seconds producing the 5 resulting files per patient. (around 2 and a half minutes producing the results for all the 3 patients). -- “NV_GPU=1” indicates the index of the GPU that will be used, without this command the method uses all available GPUs --“directory" will be your test directory containing the empty folder for your results and the files per patient that must contain the four modalities.