This project aims to use "Transfer Learning" with Keras Models to accurately classify different types of Interstitial Lung Disease (ILD). By providing CT scan patches in RGB format, the Convolutional Neural Networks can learn to identify each class of ILD, as well as those without any disease. Transfer Learning allows the program to pull intermediate outputs from the network and compare classes using simpler features.
Install miniconda.
Create a conda environment:
conda create -n Keras python=3.5 numpy scipy yaml h5py scikit-learn pillow
source activate Keras
pip install Tensorflow-gpu
pip install Keras
pip install cython
Note some of the required conda packages might be missing from this list!
Unzip images: unzip binary_images.zip unzip multiclass_images.zip
Activate the conda environment:
source activate Keras
See talisman-test-suite.py for a usage example. Also most functions have up to date docstrings on the master branch.
Dictionary of images are saved in svc file located in tl_keras_v2/research.
Put the images in folders named with the image class label. Each class needs at least 25 images. Then put these folders in a parent directory so the directory structure looks something like:
images
├── healthy
├── emphysema
├── fibrosis
├── ground_glass
└── micronodules
The research directory comes pre-loaded with 2 .csv groups files:
binary_patient_groups.csv
multiclass_patient_groups.csv
Correctly assign the groups_file variable in talisman-test-suite.py based on desired classification.