Segmentation of prostate and prostate zones using Deep Learning: a multi-MRI vendor analysis

This repository contains a test case for one of our recent projects which consist of an automatic segmentation algorithm for the prostate and its peripheral zone (PZ) using a 3D Convolutional Neural Network (CNN).

The details of our algorithm will soon be published and we will add a link in here.

With this program six different CNN models can be tested in two sample datasets. The two datasets contain MR images from different vendors: Siemens and GE. And the models are split in 3 for prostate segmentation and 3 for PZ segmentation. The difference between the models is the training dataset used to build them, not the CNN architecture.

The architecture used to train our models is the following:

Install

Important!! This code uses f-strings, which are available since Python 3.6. If you are using an older version please reformat the print statements.

Conda

conda install -c anaconda tensorflow-gpu 
conda install -c simpleitk simpleitk 
conda install -c simpleitk/label/dev simpleitk
conda install -c conda-forge keras 

Conda from requirements.txt file (linux)

while read requirement; do conda install --yes $requirement || pip install $requirement; done < requirementsT.txt

Data

The test data is freely available and can be downloaded from this link data. In order to edit the configuration file as little as possible, I suggest you to copy the data folder at the root folder of this repository.

Run Test

The organization of the folders is the one below, and its content is self explanatory. The file to run the test is inside the code folder and is MakeSegmentation.py. To configure the run you must edit the file MainConfig.py inside code/config. The configuration file is well documented and there is not need to re-explain it here. Just mention that in that file you can configure your input and output folders, which model to test, in which dataset, etc.

To test any of the models, first edit the configuration file and then run it with:

cd code
python MakeSegmentation.py

This program will make a segmentation with the proposed CNN and it will create images showing the ground truth contour and the predicted contour. It will also compute the Dice Coefficient of the segmentation and it will save it in a CSV file. Some of the images you should be able to generate with this test case are:

Prostate Segmentation	PZ Segmentation