dicommunify

DICommunify is built for dicom header images in which to translate them from imagery to the approriate DICOM encoding

Prerequisities

• virtualenv
• python 3.x installed on your computer
• scipy and numpy
• theano / tensorflow and keras
• matplotlib (optional)

Installation

• brew install pip3
• pip3 install -r requirements.txt
• mkdir -p data/raw
• download the Image_Downscaled zipfile from Forefront data repo. Ask someone who has access.
  • extract the folder to data/raw so you get data/raw/Image_Downscaled
• download the csv file which contain the image to label data at ImageData.csv
  • put the csv file into data/raw so you get data/raw/ImageData.csv

data-preprocessing

* this is to visualize the neural network inside of jupyter notebook

• open Jupyter Notebook by jupyter notebook
• go to notebooks/preprocessing.ipynb
• run through all of the cells and the output will be folders with each class
• makes it into the folder structure

train:
    Body: "images"
    Head-Neck: "images"
    ..: "images"
test:
    Body: "images"
    Head-Neck: "images"
    ..: "images"

installation of tensorflow and running any model

• for mac install brew install gprof2dot
  • otherwise install from source graphviz
• open the image_classification.ipynb notebook
• before running the model, start tensorboard by running:
  • cd to notebook directory and run
  • tensorboard --logdir summary --port=8008 &
  • this will give you tensorboard @ localhost:8008
• for running completely new models with tensorboard, just delete the summary folder
• all models will be stored inside models folder for further evaluation

TODO

[x] create dataset with samples and labels (x,y)

[x] train a model for classifying at least with > 80 % accuracy and more precision than recall