This is the folder with the starterkit to do UNET training and inference on either artificial data or on real medical images.
all work was done with Python 3.5 then Tensorflow v1.1 and opencv v3.1 the rest of the packages is best installed with anaconda or pip. No comprehensive list of packages is given here. Also, some imports may actually not be used, hence proceed with some trials.
The parameters in the main Python file are set to use the Data Reader with artifical images and corresponding labels. Hence no need for real images. The UNET generated is a 3 layer variety and should execute quite swiftly As the batch size is set to 4, enough local memory should be available for the CPU Create a directory "logs" so that events and model checkpoints can be saved in local subdirectories Best to overlook progress of training is with Tensorboard. Launch it within the "logs" directory with
tensorboard --logdir "subdirectory_name created by the script"
FCN_jul4.py : main file to launch the training.
ModelLibrary.py : defines the UNET model with the given parameters
BatchDatsetReader.py : reader for medical images prepared in the data directory (here dataJul4)
DatasetReader.py : reader for generator images
TensorflowUtils.py : generic utilities
Inference Greyscale FCN-UNET.ipynb : launches the inference on samples and visualises result
ATTENTION : THIS NOTEBOOK IS FOR THE KERAS BASED FCN
ROI processing-Jul4-stripped.ipynb : creates the data structure for medical images based on DICOM files and text files for ROIs. When done, zip the content and copy it into the data... folder
HUG_data_reader.ipynb : visualisation of data from the 2 Readers
data.... : folder for the medical image data. It expects to see at least a .zip with the images in .jpg format and the labels in .png format
logs : model checkpoint folder events folder for TensorBoard
tf_unet : folder with helper utilities for UNET model