MLMedic

Putting great tools in the hand of clinicians. https://github.com/MLMedic/MLMedic

This is the project repository for the Brisbane health hack 2019. The goal is to develop an interface for applying mashine learning models to medical imaging data.

Feature List:

1. platform-independent GUI in Python / Electron / ?
2. Import of Dicom data
3. Applying Machine Learning models to this dicom data (example: Segmentation and Highlighting of Brain Lesions)
4. Visualising Output

Optional Feature List:

• Model zoo online with upload possibility
• Model conversion from Tensorflow, PyTorch, Caffe, Theano .... to be able to be used in our GUI
• Local Transfer Learning to adjust models to available data at the local site

Getting started:

Installing node on Linux/Mac without root permissions

Install nvm from https://github.com/creationix/nvm

1. On your terminal, execute the following command.

curl -o- https://raw.githubusercontent.com/creationix/nvm/v0.33.11/install.sh | bash

2. Create the path to nvm

export NVM_DIR="${XDG_CONFIG_HOME/:-$HOME/.}nvm"
[ -s "$NVM_DIR/nvm.sh" ] && \. "$NVM_DIR/nvm.sh" # This loads nvm

3. Verify the installation

command -v nvm

4. Install your version of node using nvm

nvm install 10.10.0

5. Use your node version

nvm use 10.10.0

Installing node on Windows without root permissions

• We need add the node.exe and npm.cmd to system path, so that we can access them from any where.
• Now append the project folder spath to PATH variable in Environment Variables.
• We can access the Environment Variables dialog, by right clicking on the Computer > Properties > Advanced system settings > Advanced tab > Environment Variables.
• Incase if you don’t have permission to access Environment Variables dialog, simply type the following command in Run dialog

rundll32 sysdm.cpl,EditEnvironmentVariables

, this will open the Environment Variables dialog.

• Go to command line then type node -v then npm -v. These commands should display the currently configured nodejs and npm versions respectively.

Data:

• 3T and 7T MPRAGE and MP2RAGE anatomical scans
• dicom and nii format
• link via email

Available Models:

• https://github.com/DLTK/models/tree/master/ukbb_neuronet_brain_segmentation (Tensorflow)
• https://github.com/zsdonghao/u-net-brain-tumor.git (TensorFlow)
• https://pypi.org/project/NiftyNet/ (Nifty net - works very easily)
• https://github.com/josedolz/LiviaNET (Python 2 and Theano)
• https://github.com/Entodi/MeshNet (Torch)
• http://64.234.162.248:3000/about (Same as above)

Availabe Tools (need to be trained first):

• https://github.com/kaczmarj/nobrainer
• https://github.com/GUR9000/Deep_MRI_brain_extraction

Example how a current application of a model looks like:

https://github.com/DLTK/models/tree/master/ukbb_neuronet_brain_segmentation

• install miniconda https://conda.io/miniconda.html or anaconda
• wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
• bash Miniconda3-latest-Linux-x86_64.sh
• conda install tensorflow
• pip install dltk
• clone model repo:
  • git clone https://github.com/DLTK/models
• download Models:
• wget http://www.doc.ic.ac.uk/~mrajchl/dltk_models/model_zoo/neuronet/spm_tissue.tar.gz
• tar -xzf spm_tissue.tar.gz (into /models/ukbb_neuronet_brain_segmentation/
• copy files from spm_tissue/0/1513180449 up one level to spm_tissue/0
• adjust paths in /models/ukbb_neuronet_brain_segmentation/config_spm_tissue.json so they point to the path /models/ukbb_neuronet_brain_segmentation/spm_tissue or whatever is relevant.
• create and add this to /models/ukbb_neuronet_brain_segmentation/files.csv in two lines: id,t1,fsl_fast,fsl_first,spm_tissue,malp_em,malp_em_tissue 5404127,3T.nii.gz,T1_brain_seg.nii.gz,all_fast_firstseg.nii.gz,T1_brain_seg_spm.nii.gz,T1_MALPEM.nii.gz,T1_MALPEM_tissues.nii.gz
• download 3T file from link provided on owncloud and name it 3T.nii.gz, place it in /models/ukbb_neuronet_brain_segmentation/
• run the model!
  • python deploy.py --csv files.csv --config config_spm_tissue.json

Another example that needs Torch (if someone knows how to convert this to tensorflow/TF.js!):

From https://github.com/Entodi/MeshNet

• First you need Torch!
• Steps taken from https://dwijaybane.wordpress.com/2017/07/22/installing-torch-7-deep-learning-on-ubuntu-16-04/
  • sudo apt-get install --no-install-recommends git software-properties-common
  • export TORCH_ROOT=~/torch
  • git clone https://github.com/torch/distro.git $TORCH_ROOT --recursive
  • cd $TORCH_ROOT
  • ./install-deps
  • ./install.sh -b
• Now download the models for MeshNet AKA BrainChop
  • git clone https://github.com/Entodi/MeshNet.git
• Download the 3T data from owncloud link
• Install python and dependencies if you haven't:
  • pip install nipy
• Conform T1 to 1mm voxel size in coronal slice direction with side length 256.
  • (Freesurfer required) mri_convert brainDir/t1.nii brainDir/t1_c.nii -c
• Convert nifti to numpy format
  • python nifti2npy.py brainDir/t1_c.nii --npy_file brainDir/T1.npy
• Create segmentation using predict.lua providing path to directory with brain npy file brainDir
  • th predict.lua -modelFile ./saved_models/model_Mon_Jul_10_16:43:55_2017/model_219.t7 -brainPath brainDir
• Convert numpy segmentation file to nifti format by providing base nifti file
  • python npy2nifti.py segmentation.npy t1_c.nii

Can we replace this with a nice GUI that ideally doesnt need a python installation?

• Electron
• https://www.npmjs.com/package/dicom
• https://www.npmjs.com/package/imageviewer
• TF.js (https://js.tensorflow.org/)
• https://github.com/bioimagesuiteweb/bisweb

Data to play with

https://cloudstor.aarnet.edu.au/plus/s/JGt804o3cGXc5xf

Dicom to nii converter in python3:

pip install git+git://github.com/mvdoc/dcmstack@bf/importsys