methods.tex

%  pdflatex methods.tex; bibtex methods; pdflatex methods.tex; pdflatex methods.tex
%  This simple example illustrates how documents can be
%  split into smaller segments, each segment processed
%  by latex2html separately.  This document can be
%  processed through latex and latex2html with the
%  corresponding makefile.
%

\documentclass{article}         % Must use LaTeX 2e
\usepackage[plainpages=false, colorlinks=true, citecolor=black, filecolor=black, linkcolor=black, urlcolor=black]{hyperref}		
\usepackage[left=.75in,right=.75in,top=.75in,bottom=.75in]{geometry}
\usepackage{makeidx,color,boxedminipage}
\usepackage{graphicx,float}
\usepackage{amsmath,amsthm,amsfonts,amscd,amssymb} 

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%	Some math support.					     %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%	Theorem environments (these need the amsthm package)
%
%% \theoremstyle{plain} %% This is the default

\newtheorem{thm}{Theorem}[section]
\newtheorem{cor}[thm]{Corollary}
\newtheorem{lem}[thm]{Lemma}
\newtheorem{prop}[thm]{Proposition}
\newtheorem{ax}{Axiom}

\theoremstyle{definition}
\newtheorem{defn}{Definition}[section]

\theoremstyle{remark}
\newtheorem{rem}{Remark}[section]
\newtheorem*{notation}{Notation}
\newtheorem*{exrcs}{Exercise}
\newtheorem*{exmple}{Example}

%\numberwithin{equation}{section}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%	Macros.							     %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%	Here some macros that are needed in this document:

\newcommand{\motion}{\mathbf{\varphi}}
\newcommand{\hmotion}{\mbox{\boldmath $\hat{\varphi}$}}
\newcommand{\cauchy}{\mbox{\boldmath $\sigma$}}
\newcommand{\eqn}[1]{(\ref{#1})}
\newcommand{\hOmega}{\hat{\Omega}}
\newcommand{\homega}{\hat{\omega}}
\newcommand{\nphalf}{n+\frac{1}{2}}
\newcommand{\nmhalf}{n-\frac{1}{2}}
\newcommand{\kmhalf}{k-\frac{1}{2}}
\newcommand{\kphalf}{k+\frac{1}{2}}
\newcommand{\picdir}{pdffig/}

\include{symbols}
\title{Methods}
\author{}
\begin{document}                % The start of the document
\maketitle

\textbf{MR imaging data}.  Imaging was performed on 1.5T MR scanners:
XXX scanner, GE Healthcare, Waukesha, WI, USA , YYY scanner  Siemens, and ZZZ scanner, Philips.
Preoperative abdominal MR was obtained using a  liver protocol comprised of a
precontrast phase, a venous phase (60-80 seconds after injection of intravenous contrast
material), and a delayed phase (15 minutes after contrast injection).
Each series was obtained in the axial plane with a
phased‐array multicoil.
An T1‐weighted spoiled gradient‐echo sequence 
was performed (TR/TE = 4.6ms/2.3ms, flip angle = 15deg, breath‐hold = 20 sec,
section thickness = 3 mm, intersection gap = 1.5 mm, field‐of‐view (FOV) = 36x40cm, and
acquisition matrix, 320x255).
{\color{red} @newsha - need BCM imaging protocol and scanner hardware/software version details here.}

The multi-phase MR studies (obtained 60-80 seconds after contrast medium
administration) were exported in DICOM format from the picture archiving and communication
system to an independent server running Slicer~\cite{kikinis20143d}.
The slicer scripting interface was use to iteratively anonymize all data and maintain
orientation and resolution information as a compressed
NifTI format recommended by the Neuroimaging Informatics Technology Initiative.



\textbf{Image normalization}.
Image intensity harmonization was achieved in two steps: bias correction followed by z-score normalization. 
All protocol phases were considered independently.
N4 bias correction~\cite{tustison2010n4itk} was applied  with  4 levels and a factor of two decrease in resolution
at each level. 
MR images where shifted by 1 to avoid logarithm errors for background pixels with intensity value equal 0.
Z-score normalization was achieved by subtracting the volume-wise mean of the image and dividing by the 
standard deviation of the image volume.
Convert3D image-processing tool~\cite{yushkevich2006user} was used for image normalization.

\textbf{Image registration}.
All longitudinal data was deformably registered to a common time point for analysis.
Cases were deformably registered to the diagnostic scan.
Controls  were deformably registered to the first available scan.
A symmetric normalization (SyN) approach \cite{fuentes2015morphometry}  is a popular
variant of diffeomorphic image registration and is used in this study.
SyN calculates image gradients only at the
midpoint in time of the full diffeomorphic transformation and provides an explicit
symmetry of the large deformation diffeomorphic metric mapping formulation~\cite{beg2005computing}.
The registration consisted of a rigid registration step, an
affine registration step, a diffeomorphic step. Multi-resolution is applied in all steps.
The default gradient descent was used in
the optimization of the affine and diffeomorphic registration. A Gaussian regularization
kernel with an isotropic width of 3 voxels was used as an estimate of the Green's kernel
for the deformation operator.
Liver masks were used at all resolutions for both the fixed and moving image of the deformable image registration.


\textbf{Liver mask}.
A 3D neural network was used to generate liver masks to facilitate efficient deformable image
registration. Our previous neural network implementations~\cite{Gates2018,Morshid2019} in
computed tomography were adapted to MR data.  Briefly, network architecture follows a
hybrid DenseNet and U-Net architecture~\cite{huang2017densely,ronneberger2015u,kamnitsas2016deepmedic}.  The
architecture is constructed from a composition of convolution and down sampling operations
that extract features along a contracting path.  Similarly, an expanding path consists of
convolution and up-sampling operations with 'long skip' connections to integrate features
from the corresponding down sampling operations. Four resolution levels are used.  At a
given resolution, the feature-maps of all preceding layers are used as inputs, and its own
feature-maps are used as inputs into all subsequent layers.  Each convolutional operation
uses a 3x3 kernel size and is followed by a batch normalization and a ReLU activation function.

An independent training set of N=34 multi-contrast MR image data sets were registered and
labeled in Velocity.
Each data set consisted of a pre-contrast, venous phase, and delayed phase image. A manual
liver mask for each phase was created using semi-automated segmentation tools in velocity.
{\color{red}(@matt - need a few sentences here - what was your labeling protocol ?)}


Training was contrast agnostic \cite{thakur2020brain} and applied to all 102 images in the
training set. The neural network was trained on 64x64x64 patches with a batch size of
eight for both training and validation data. Sixteen random patches were selected  per
image.  The Adam optimizer was used for 50 epochs with an intial learning rate of 5e-4 and
learning rate drop factor of 0.95 every 5 epochs. Evaluation of the validation set was
performed with a frequency of 180 iterations.

Notably, although the patch size for training is 64\textsuperscript{3},
convolution, ReLU, and Normalization layers are agnostic to image size. 
Test set predictions are applied to the full data resolution.
Linear resampling to 256x256 pixels in-plane resolution is needed to fit the entire image in the GPU RAM
and CUDA intmax('int32') overflow restrictions.
Image data  was resized to 256x256 pixels in-plane resolution prior to patch extraction for
training as well as prior to the test set predictions.
Segmentation predictions are resized back the original resolution.
Liver mask predictions were dilated by 15voxels in all directions and used an a ROI for the deformable image registration
All training and test set predictions were performed on an Nvidia Quadro RTX 8000 card with 4608 CUDA cores and 48GB RAM.

\textbf{Discussion}.  Tas, what was our wishlist EPM improvements? objective selection of
liver ROI? can you put this in the discussion please? 


\bibliographystyle{apalike}
\bibliography{references}

\begin{verbatim}
maeda$ make -n -B bcmdata/BCM0025010/Pst.raw.nii.gz
mkdir -p bcmdata/BCM0025010; c3d  LiverMRIProjectData/tmpconvert//BCM0025010/c9e93ebe-cc0a-fa36-6ca1-250d842f9c65.nii.gz  -o bcmdata/BCM0025010/Pst.raw.nii.gz

python normalization.py --imagefile=bcmdata/BCM0025011/fixed.raw.nii.gz  --output=bcmdata/BCM0025011/fixed.zscore.nii.gz
python resize.py --imagefile=bcmdata/BCM0025011/fixed.zscore.nii.gz  --output=bcmdata/BCM0025011/fixed.crop.nii.gz

maeda$ make -n -B  Processed/hccmrilog/dscimg/densenet3d/adadelta/256/hccmrima/005020/001/000/trainedNet.mat
matlab -nodesktop -softwareopengl -r "livermodel('Processed/hccmrilog/dscimg/densenet3d/adadelta/256/hccmrima/005020/001/000/setup.json');exit"
maeda$ make -n -B bcmdata/BCM0025010/Pst.label.nii.gz
echo applymodel\('bcmdata/BCM0025010/Pst.256.nii.gz','Processed/hccmrilog/dscimg/densenet3d/adadelta/256/hccmrima/005020/001/000/restore_10162020/trainedNet.mat','bcmdata/BCM0025010/Pst','1','gpu'\)
mkdir -p bcmdata/BCM0025010/Pst;./run_applymodel.sh /opt/apps/matlab/R2020a/ bcmdata/BCM0025010/Pst.256.nii.gz Processed/hccmrilog/dscimg/densenet3d/adadelta/256/hccmrima/005020/001/000/restore_10162020/trainedNet.mat bcmdata/BCM0025010/Pst 1 gpu
echo vglrun itksnap -g bcmdata/BCM0025010/Pst.256.nii.gz -s bcmdata/BCM0025010/Pst/label.nii.gz -o bcmdata/BCM0025010/Pst/score.nii.gz
c3d -verbose bcmdata/BCM0025010/Pst.raw.nii.gz bcmdata/BCM0025010/Pst/label.nii.gz -reslice-identity -o bcmdata/BCM0025010/Pst.label.nii.gz
maeda$ make -n -B bcmdata/BCM0025010/Pst.regcc.nii.gz
c3d -verbose bcmdata/BCM0025010/Pst.raw.nii.gz  -shift 1  -o  bcmdata/BCM0025010/Pst.bias.nii.gz
/opt/apps/ANTS/dev/install/bin/N4BiasFieldCorrection -v 1 -d 3 -c [20x20x20x10,0] -b [200] -s 2 -i  bcmdata/BCM0025010/Pst.bias.nii.gz  -o  bcmdata/BCM0025010/Pst.bias.nii.gz
python normalization.py --imagefile=bcmdata/BCM0025010/Pst.bias.nii.gz  --output=bcmdata/BCM0025010/Pst.bias.nii.gz
c3d -verbose bcmdata/BCM0025010/Pst.label.nii.gz  -thresh 2 2 1 0  -comp -thresh 1 1 1 0  -o  bcmdata/BCM0025010/Pst.liver.nii.gz -dilate 1 15x15x15vox -o bcmdata/BCM0025010/Pst.mask.nii.gz
export ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS=24; /opt/apps/ANTS/dev/install/bin/antsRegistration --verbose 1 --dimensionality 3 --float 0 --collapse-output-transforms 1 --output [bcmdata/BCM0025010/Pst.regcc,bcmdata/BCM0025010/Pst.regcc.nii.gz] --interpolation Linear --use-histogram-matching 0 --winsorize-image-intensities [ 0.005,0.995 ] -x [bcmdata/BCM0025010/Art.mask.nii.gz,bcmdata/BCM0025010/Pst.mask.nii.gz] --transform Rigid[ 0.1 ] --metric MI[ bcmdata/BCM0025010/Art.bias.nii.gz,bcmdata/BCM0025010/Pst.bias.nii.gz,1,32,Regular,0.25 ] --convergence [ 1000x500x250x100,1e-6,10 ] --shrink-factors 8x4x2x1 --smoothing-sigmas 3x2x1x0vox --transform Affine[ 0.1 ] --metric MI[ bcmdata/BCM0025010/Art.bias.nii.gz,bcmdata/BCM0025010/Pst.bias.nii.gz,1,32,Regular,0.25 ] --convergence [ 1000x500x250x100,1e-6,10 ] --shrink-factors 8x4x2x1 --smoothing-sigmas 3x2x1x0vox --transform SyN[ 0.1,3,0 ] --metric CC[ bcmdata/BCM0025010/Art.bias.nii.gz,bcmdata/BCM0025010/Pst.bias.nii.gz,1,4 ] --convergence [
100x70x50x20,1e-6,10 ] --shrink-factors 8x4x2x1 --smoothing-sigmas 3x2x1x0vox > bcmdata/BCM0025010/Pst.regcc.log  2>&1



maeda$ cat Processed/hccmrilog/dscimg/densenet3d/adadelta/256/hccmrima/005020/001/000/setup.json
{"stoFoldername": "hccmrilog", "testset": null, "nnmodel": "densenet3d", "validationset": ["prehcc0031/scaled", "prehcc0031/bias", "prehcc0031/zscore", "arthcc0031/scaled", "arthcc0031/bias", "arthcc0031/zscore", "venhcc0031/scaled", "venhcc0031/bias", "venhcc0031/zscore", "prehcc0032/scaled", "prehcc0032/bias", "prehcc0032/zscore", "arthcc0032/scaled", "arthcc0032/bias", "arthcc0032/zscore", "venhcc0032/scaled", "venhcc0032/bias", "venhcc0032/zscore", "prehcc0033/scaled", "prehcc0033/bias", "prehcc0033/zscore", "arthcc0033/scaled", "arthcc0033/bias", "arthcc0033/zscore", "venhcc0033/scaled", "venhcc0033/bias", "venhcc0033/zscore", "prehcc0034/scaled", "prehcc0034/bias", "prehcc0034/zscore", "arthcc0034/scaled", "arthcc0034/bias", "arthcc0034/zscore", "venhcc0034/scaled", "venhcc0034/bias", "venhcc0034/zscore"], "kfold": 35, "NumberOfChannels": 1, "uidoutputdir": "Processed/hccmrilog/dscimg/densenet3d/adadelta/256/hccmrima/005020/001/000", "resolution": 256, "trainset":
["prehcc0001/scaled", "prehcc0001/bias", "prehcc0001/zscore", "arthcc0001/scaled", "arthcc0001/bias", "arthcc0001/zscore", "venhcc0001/scaled", "venhcc0001/bias", "venhcc0001/zscore", "prehcc0002/scaled", "prehcc0002/bias", "prehcc0002/zscore", "arthcc0002/scaled", "arthcc0002/bias", "arthcc0002/zscore", "venhcc0002/scaled", "venhcc0002/bias", "venhcc0002/zscore", "prehcc0003/scaled", "prehcc0003/bias", "prehcc0003/zscore", "arthcc0003/scaled", "arthcc0003/bias", "arthcc0003/zscore", "venhcc0003/scaled", "venhcc0003/bias", "venhcc0003/zscore", "prehcc0004/scaled", "prehcc0004/bias", "prehcc0004/zscore", "arthcc0004/scaled", "arthcc0004/bias", "arthcc0004/zscore", "venhcc0004/scaled", "venhcc0004/bias", "venhcc0004/zscore", "prehcc0005/scaled", "prehcc0005/bias", "prehcc0005/zscore", "arthcc0005/scaled", "arthcc0005/bias", "arthcc0005/zscore", "venhcc0005/scaled", "venhcc0005/bias", "venhcc0005/zscore", "prehcc0006/scaled", "prehcc0006/bias", "prehcc0006/zscore", "arthcc0006/scaled",
"arthcc0006/bias", "arthcc0006/zscore", "venhcc0006/scaled", "venhcc0006/bias", "venhcc0006/zscore", "prehcc0007/scaled", "prehcc0007/bias", "prehcc0007/zscore", "arthcc0007/scaled", "arthcc0007/bias", "arthcc0007/zscore", "venhcc0007/scaled", "venhcc0007/bias", "venhcc0007/zscore", "prehcc0008/scaled", "prehcc0008/bias", "prehcc0008/zscore", "arthcc0008/scaled", "arthcc0008/bias", "arthcc0008/zscore", "venhcc0008/scaled", "venhcc0008/bias", "venhcc0008/zscore", "prehcc0009/scaled", "prehcc0009/bias", "prehcc0009/zscore", "arthcc0009/scaled", "arthcc0009/bias", "arthcc0009/zscore", "venhcc0009/scaled", "venhcc0009/bias", "venhcc0009/zscore", "prehcc0010/scaled", "prehcc0010/bias", "prehcc0010/zscore", "arthcc0010/scaled", "arthcc0010/bias", "arthcc0010/zscore", "venhcc0010/scaled", "venhcc0010/bias", "venhcc0010/zscore", "prehcc0011/scaled", "prehcc0011/bias", "prehcc0011/zscore", "arthcc0011/scaled", "arthcc0011/bias", "arthcc0011/zscore", "venhcc0011/scaled", "venhcc0011/bias",
"venhcc0011/zscore", "prehcc0012/scaled", "prehcc0012/bias", "prehcc0012/zscore", "arthcc0012/scaled", "arthcc0012/bias", "arthcc0012/zscore", "venhcc0012/scaled", "venhcc0012/bias", "venhcc0012/zscore", "prehcc0013/scaled", "prehcc0013/bias", "prehcc0013/zscore", "arthcc0013/scaled", "arthcc0013/bias", "arthcc0013/zscore", "venhcc0013/scaled", "venhcc0013/bias", "venhcc0013/zscore", "prehcc0014/scaled", "prehcc0014/bias", "prehcc0014/zscore", "arthcc0014/scaled", "arthcc0014/bias", "arthcc0014/zscore", "venhcc0014/scaled", "venhcc0014/bias", "venhcc0014/zscore", "prehcc0015/scaled", "prehcc0015/bias", "prehcc0015/zscore", "arthcc0015/scaled", "arthcc0015/bias", "arthcc0015/zscore", "venhcc0015/scaled", "venhcc0015/bias", "venhcc0015/zscore", "prehcc0016/scaled", "prehcc0016/bias", "prehcc0016/zscore", "arthcc0016/scaled", "arthcc0016/bias", "arthcc0016/zscore", "venhcc0016/scaled", "venhcc0016/bias", "venhcc0016/zscore", "prehcc0017/scaled", "prehcc0017/bias", "prehcc0017/zscore",
"arthcc0017/scaled", "arthcc0017/bias", "arthcc0017/zscore", "venhcc0017/scaled", "venhcc0017/bias", "venhcc0017/zscore", "prehcc0018/scaled", "prehcc0018/bias", "prehcc0018/zscore", "arthcc0018/scaled", "arthcc0018/bias", "arthcc0018/zscore", "venhcc0018/scaled", "venhcc0018/bias", "venhcc0018/zscore", "prehcc0019/scaled", "prehcc0019/bias", "prehcc0019/zscore", "arthcc0019/scaled", "arthcc0019/bias", "arthcc0019/zscore", "venhcc0019/scaled", "venhcc0019/bias", "venhcc0019/zscore", "prehcc0020/scaled", "prehcc0020/bias", "prehcc0020/zscore", "arthcc0020/scaled", "arthcc0020/bias", "arthcc0020/zscore", "venhcc0020/scaled", "venhcc0020/bias", "venhcc0020/zscore", "prehcc0021/scaled", "prehcc0021/bias", "prehcc0021/zscore", "arthcc0021/scaled", "arthcc0021/bias", "arthcc0021/zscore", "venhcc0021/scaled", "venhcc0021/bias", "venhcc0021/zscore", "prehcc0022/scaled", "prehcc0022/bias", "prehcc0022/zscore", "arthcc0022/scaled", "arthcc0022/bias", "arthcc0022/zscore", "venhcc0022/scaled",
"venhcc0022/bias", "venhcc0022/zscore", "prehcc0023/scaled", "prehcc0023/bias", "prehcc0023/zscore", "arthcc0023/scaled", "arthcc0023/bias", "arthcc0023/zscore", "venhcc0023/scaled", "venhcc0023/bias", "venhcc0023/zscore", "prehcc0024/scaled", "prehcc0024/bias", "prehcc0024/zscore", "arthcc0024/scaled", "arthcc0024/bias", "arthcc0024/zscore", "venhcc0024/scaled", "venhcc0024/bias", "venhcc0024/zscore", "prehcc0025/scaled", "prehcc0025/bias", "prehcc0025/zscore", "arthcc0025/scaled", "arthcc0025/bias", "arthcc0025/zscore", "venhcc0025/scaled", "venhcc0025/bias", "venhcc0025/zscore", "prehcc0026/scaled", "prehcc0026/bias", "prehcc0026/zscore", "arthcc0026/scaled", "arthcc0026/bias", "arthcc0026/zscore", "venhcc0026/scaled", "venhcc0026/bias", "venhcc0026/zscore", "prehcc0027/scaled", "prehcc0027/bias", "prehcc0027/zscore", "arthcc0027/scaled", "arthcc0027/bias", "arthcc0027/zscore", "venhcc0027/scaled", "venhcc0027/bias", "venhcc0027/zscore", "prehcc0028/scaled", "prehcc0028/bias",
"prehcc0028/zscore", "arthcc0028/scaled", "arthcc0028/bias", "arthcc0028/zscore", "venhcc0028/scaled", "venhcc0028/bias", "venhcc0028/zscore", "prehcc0029/scaled", "prehcc0029/bias", "prehcc0029/zscore", "arthcc0029/scaled", "arthcc0029/bias", "arthcc0029/zscore", "venhcc0029/scaled", "venhcc0029/bias", "venhcc0029/zscore", "prehcc0030/scaled", "prehcc0030/bias", "prehcc0030/zscore", "arthcc0030/scaled", "arthcc0030/bias", "arthcc0030/zscore", "venhcc0030/scaled", "venhcc0030/bias", "venhcc0030/zscore"]}
\end{verbatim}
\end{document}