Awesome

Fast motion compensation and super-resolution from multiple stacks of 2D slices

PVR - Patch to Volume Reconstruction

PVR is a research-focused tool for the task of superresolution image reconstruction, developed at the BioMedIA research group. It provides command-line tools to reconstruct images from motion-corrupted imaging data.

Dependencies

• Boost = v1.58.0
• CUDA >= v 6.5.0
• Intel TBB

Installation

$ cd <path-to-cudarecon/source> 
$ mkdir build
$ cd build
$ cmake ..
$ make 

Usage

• PVR reconstruction using square patches

$ PVRreconstructionGPU -i <path-to-input-data> -o <reconstructed-image-filename> --resolution <1> --patchSize <x> <y> --patchStride <x> <y>

• PVR reconstruction using superpixels

$ PVRreconstructionGPU -i <path-to-input-data> -o <reconstructed-image-filename> --resolution <1> --superpixel --spxSize <x> --spxExtend <y>

• PVR reconstruction using square patches

$ SVRreconstructionGPU -i <path-to-input-data> -o <reconstructed-image-filename> --resolution <1> 

Docker

thanks to <a href="https://github.com/dittothat">Jeff Stout</a> there is a docker container available. from <a href="https://github.com/dittothat">dittohat</a>'s post:

<a href="https://github.com/dittothat">Docker image</a>

<a href="https://github.com/dittothat/dockerfetalrecon/blob/master/Dockerfile">Dockerfile</a>

On a system with nvidia-docker installed (Docker 19.03 only requires that you use the --gpus option), do:

docker pull dittothat/fetalreconstruction:cuda6.5

Then you can run using:

docker run --gpus all -it --mount type=bind,source=<host path to data>,target=/data dittothat/fetalreconstruction:cuda6.5 /bin/bash

Inside the container:

cd /usr/src/fetalReconstruction/data PVRreconstructionGPU -o 3TReconstruction.nii.gz -i 14_3T_nody_001.nii.gz 10_3T_nody_001.nii.gz 21_3T_nody_001.nii.gz 23_3T_nody_001.nii.gz -m mask_10_3T_brain_smooth.nii.gz --resolution 1.0

Data you wish to process can be put in whatever path you specify, and is then located at /data/ in the container.

TODO

• the maximum patch size is currently limited by the abilities of the used GPU. A maximum size of 64x64 is safe for average high end cards. This issue will be fixed soon.

References

• [arxiv16] "PVR: Patch-to-Volume Reconstruction for Large Area Motion Correction of Fetal MRI." Amir Alansary, Bernhard Kainz, Martin Rajchl, Maria Murgasova, Mellisa Damodaram, David F.A. Lloyd, Steven G. McDonagh, Alice Davidson, Mary Rutherford, Reza Razavi, Joseph V. Hajnal, and Daniel Rueckert.
• [MICCAI15] "Flexible reconstruction and correction of unpredictable motion from stacks of 2D images." Bernhard Kainz, Amir Alansary, Christina Malamateniou, Kevin Keraudren, Mary Rutherford, Joseph V. Hajnal, and Daniel Rueckert. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 555-562. Springer International Publishing, 2015.
• [MEDIA12] "Reconstruction of fetal brain MRI with intensity matching and complete outlier removal." Maria Kuklisova-Murgasova, Gerardine Quaghebeur, Mary A. Rutherford, Joseph V. Hajnal, and Julia A. Schnabel. Medical image analysis 16, no. 8 (2012): 1550-1564.

License

PVR is licensed under the MIT license

The following are compilation notes by Shadab Khan skhan.shadab@gmail.com for MacOS:

1. Copied zlib.h and zconf.h from zlib-1.2.5 (which I downloaded from sourceforge) and pasted inside irtk2/nifti/zlib/. Renamed original zlib to zlib_old -> This issue occurs because OS X uses zlib version 1.2.5 internally (located in /usr/lib and /usr/include) which cannot be modified by the user. Version 1.2.5 (used by OS X) doesn't contain all functions in 1.2.7 (shipped with fetal-reconstruction source), hence the problem occurs. This is a quick fix, I am forcing functionality available in zlib 1.2.5 which doesn't seem to affect any of file i/o. I should find proper CMake commands to explicitly link against user defined zlib.

2. fetalReconstruction-master/source/IRTKSimple2/geometry++/include/irtkMatrix.h ERROR: friend declaration specifying a default argument must be a definition Commented line 175: friend irtkMatrix FrechetMean(irtkMatrix *, int, int = 10); Commented line 177: friend irtkMatrix FrechetMean(irtkMatrix *, double *, int, int = 10);

This in turn required commenting two other lines (see below)

3. fetalReconstruction-master/source/IRTKSimple2/packages/transformation/src/irtkTemporalHomogeneousTransformation.cc Commented line 59: _transformMatrix = FrechetMean(M, weight, 4); Commented line 119: _transformMatrix = FrechetMean(M, weight, 2);

4. helper_cuda.h not found: Copied headers from CUDA_DIR/Samples/common/inc/ to CUDA_DIR/include/

5. Added following to CMakeLists.txt inside /source: set(CMAKE_CXX_FLAGS "-Wno-c++11-narrowing")

SVR - Slice to Volume Reconstruction

SVR was used to produce the results shown in

Bernhard Kainz, Markus Steinberger, Wolfgang Wein, Maria Kuklisova-Murgasova, Christina Malamateniou, Kevin Keraudren, Thomas Torsney-Weir, Mary Rutherford, Paul Aljabar, Joseph V. Hajnal, and Daniel Rueckert: Fast Volume Reconstruction from Motion Corrupted Stacks of 2D Slices. IEEE Transactions on Medical Imaging, in print, 2015. doi:10.1109/TMI.2015.2415453

Announcements

• check out Michael Ebner's work improving upon our work with his slower but more recent and more robust framework NiftyMIC

• The --useCPU flag has some bug currently and might have caused some confusion. CPU only does not provide the full functionality of the GPU version. For example PSF sampling is different and the result of the CPU version will be worse. -- I am working on it.

• Hacked a fix for the working CPU version -- please report if you experience more bugs with that

• There seem to be an issue with CUDA 7.0. Please use CUDA 6.5 until I found the problem.

• The code requires refactoring to cleanly separate data from data access and CPU and GPU parts.

Abstract

Moving objects cause motion artefacts when their enclosing volume is acquired as a stack of image slices. In this paper we present a fast multi-GPU accelerated implementation of slice-to-volume registration based super-resolution reconstruction with automatic outlier rejection and intensity bias correction. We introduce a novel fully automatic selection procedure for the image stack with the least motion, which will serve as an initial registration template. We fully evaluate our method and its high dimensional parameter space. Testing is done using artificially motion corrupted phantom datasets and using real world scenarios for the reconstruction of foetal organs from in-utero prenatal Magnetic Resonance Imaging and for the motion compensation of freehand compound Ultrasound data. On average we achieve a speed-up of more than 40x compared to a single CPU system, and another 1.70x for each additional GPU, while maintaining the same image quality as if calculated on a CPU. Our framework is qualitatively more accurate and on average $10\times$ times faster than currently available state-of-the-art multi-core methods. The source code for this approach is open source and publicly available for download.

Installation

The source code was successfully compiled on x86_64, CUDA 6.5, VS2012 (std-c++11), VS2013 (boost), and gcc 4.8. System: Intel Xeon E5-2630 v2 @ 2.60GHz system with 16 GB RAM, an Nvidia Tesla K40 with 12 GB RAM and a Geforce 780 Graphics card with 6 GB RAM and also tested on a Nvidia Titan GPU.

Necessary Hardware requirements

• x86_64 CPU

Optional Hardware requirements

• Nvidia GPU compute capability >= 3.0 In case of no GPU: only CPU reconstruction will be available. Use --useCPU flag [CUDA is still necessary to compile!] HOWEVER: The CPU version only allows a simplified PSF evaluation (only linearly low sampled Gaussian PSF)! ONLY THE GPU ACCELERATED VERSION WILL PROVIDE ALL IN THE PAPER DESCRIBED FEATURES!

Necessary third party libraries

• Nvidia CUDA >= 6.0 https://developer.nvidia.com/cuda-downloads
• Boost (compiled libraries, we used 1.55.0) http://www.boost.org/users/download/
• Intel's TBB https://www.threadingbuildingblocks.org/

Optional third party libraries

• CULA http://www.culatools.com/
• IRTK http://www.doc.ic.ac.uk/~dr/software/ (This is already included in the source code repository as minimal build lib. We only use image handling and CPU registration functionality from IRTK)

BUILD INSTRUCTIONS

• go to the source directory
• make a new directory "build"
• execute cmake .. in this directory (give paths to TBB and other third party libs if necessary using cmakegui .. or ccmake ..)
• make or open the Visual Studio Solution

Usage

• get two or more overlapping stacks of image slices

• [optional]: generate a region of interest and sace as .nii or .nii.gz

• run the program SVRreconstructionGPU (previously reconstruction_GPU2)

• Example: ./SVRreconstructionGPU -o 3T_GPUtest.nii -i ../../data/14_3T_nody_001.nii.gz ../../data/10_3T_nody_001.nii.gz ../../data/23_3T_nody_001.nii.gz ../../data/21_3T_nody_001.nii.gz -m ../../data/mask_10_3T_brain_smooth.nii.gz

• Options:

  -h [ --help ] Print usage messages

  -o [ --output ] arg Name for the reconstructed volume. Nifti or Analyze format.

  -m [ --mask ] arg Binary mask to define the region od interest. Nifti or Analyze format.

  -i [ --input ] arg [stack_1] .. [stack_N] The input stacks. Nifti or Analyze format.

  -t [ --transformation ] arg The transformations of the input stack to template in 'dof' format used in IRTK. Only rough alignment with correct orienation and some overlap is needed. Use 'id' for an identity transformation for at least one stack. The first stack with 'id' transformation will be resampled as template.

  --thickness arg [th_1] .. [th_N] Give slice thickness.[Default: twice voxel size in z direction]

  -p [ --packages ] arg Give number of packages used during acquisition for each stack. The stacks will be split into packages during registration iteration 1 and then into odd and even slices within each package during registration iteration 2. The method will then continue with slice to volume approach. [Default: slice to volume registration only]

  --iterations arg (=4) Number of registration-reconstruction iterations.

  --sigma arg (=12) Stdev for bias field. [Default: 12mm] --resolution arg (=0.75) Isotropic resolution of the volume. [Default: 0.75mm]

  --multires arg (=3) Multiresolution smooting with given number of levels. [Default: 3]

  --average arg (=700) Average intensity value for stacks [Default: 700]

  --delta arg (=150) Parameter to define what is an edge. [Default: 150]

  --lambda arg (=0.02) Smoothing parameter. [Default: 0.02]

  --lastIterLambda arg (=0.01) Smoothing parameter for last iteration. [Default: 0.01]

  --smooth_mask arg (=4) Smooth the mask to reduce artefacts of manual segmentation. [Default: 4mm]

  --global_bias_correction arg (=0) Correct the bias in reconstructed image against previous estimation.

  --low_intensity_cutoff arg (=0.01) Lower intensity threshold for inclusion of voxels in global bias correction.

  --force_exclude arg force_exclude [number of slices] [ind1] ... [indN] Force exclusion of slices with these indices.

  --no_intensity_matching arg Switch off intensity matching.

  --log_prefix arg Prefix for the log file.

  --debug arg (=0) Debug mode - save intermediate results.

  --debug_gpu Debug only GPU results.

  --rec_iterations_first arg (=4) Set number of superresolution iterations

  --rec_iterations_last arg (=13) Set number of superresolution iterations for the last iteration

  --num_stacks_tuner arg (=0) Set number of input stacks that are really used (for tuner evaluation, use only first x)

  --no_log arg (=0) Do not redirect cout and cerr to log files.

  -d [ --devices ] arg Select the CP > 3.0 GPUs on which the reconstruction should be executed. Default: all devices > CP 3.0

  --tfolder arg [folder] Use existing slice-to-volume transformations to initialize the reconstruction.

  --sfolder arg [folder] Use existing registered slices and replace loaded ones (have to be equally many as loaded from stacks).

  --referenceVolume arg Name for an optional reference volume. Will be used as inital reconstruction.

  --T1PackageSize arg is a test if you can register T1 to T2 using NMI and only one iteration

  --useCPU use CPU for reconstruction and registration; performs superresolution and robust statistics on CPU. Default is using the GPU

  --useCPUReg use CPU for more flexible CPU registration; performs superresolution and robust statistics on GPU. [default, best result]

  --useGPUReg use faster but less accurate and flexible GPU registration; performs superresolution and robust statistics on GPU.

  --useAutoTemplate select 3D registration template stack automatically with matrix rank method.

  --disableBiasCorrection disable bias field correction for cases with little or no bias field inhomogenities (makes it faster but less reliable for stron intensity bias)

License

PVR is licensed under the MIT license