%% Plos One Journal - PONE-D-16-29044
%
% ##################################
% ''Multi-channel MRI segmentation of eye structures and tumors using 
%  patient-specific features''
% ##################################
%
% C. Ciller, S.I. De Zanet, K. Kamnitsas, P. Maeder, B. Glocker, 
% F.L. Munier, D. Rueckert, J-P. Thiran, M.B. Cuadra* and R. Sznitman*

*Equally contributed authors

% Copyright (c) - All rights reserved. University of Lausanne. 2016.

The content of these folders include all the necessary steps for compu-
ting the automatic segmentation of eye structures and tumors in 3D MRI.
Upon acceptance of this manuscript, all the experiments and a working
copy of the software will be made available for its use.

% Requirements:
%%%%%%%%%%%%%%%%%%%%%
Software:
- VirtualBox
- Matlab R2014a and superior
- Mac OS X / Linux / Windows

Hardware:
- 25 GB Free Disk (Virtual Image)
- 4/8 GB of RAM
- Nvidia (R) GPU (GTX 970 or superior) with CUDA/cuDNN capabilities.

% 1_EyeSegmentation :
%%%%%%%%%%%%%%%%%%%%%

admin_password: plosone

This section Covers the EyeModeler software for the segmentation of eye
structures in 3D MRI. It contains a working copy of the software deve-
loped for the automatic segmentation of eye structures in 3D MRI.

For the sake of simplicity we offer a video representing the process of
segmentation https://youtu.be/0n5Wz8RPQ7w

% 2_PlosOne_RF_Experiments :
%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Here we cover the first round of experiments with a Random Forest confi-
guration using the output of the previous step.

% 3_PlosOne_CNN_Experiments :
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

This section covers the procedure for training the DeepMedic CNN model.
The original code can be found in: 

https://github.com/Kamnitsask/deepmedic

The code in this section is modified to be able to cope with a varying
number of input channels [6,9] and a leave-one-out training configuration

% 4_EyeTumorSegment_Graphcut :
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Here we perform the final graph-cut refinement on the output of the 
experiments with different classifiers.

For the sake of simplicity, we have configured the graph-cut optimization
on the Random Forest experiments proposed in this manuscript.