RGDSVR

Tools for Robust Generative Diffeomorphic Slice to Volume Reconstructions (RGDSVR)

This repository provides tools to implement the methods in the manuscript ''Fetal MRI by robust deep generative prior reconstruction and diffeomorphic registration: application to gestational age prediction'', L Cordero-Grande, JE Ortuño-Fisac, A Uus, M Deprez, A Santos, JV Hajnal, and MJ Ledesma-Carbayo, arXiv, 2021.

The code has been developed in MATLAB and has the following structure:

./

contains a script to run a reconstruction of the provided example data: rgdsvr_example.m and another to import the Python code loadPythonDeepFetal.m.

./SVR

contains files to perform SVR reconstructions: svrAlternateMinimization.m, svrCG.m, svrDD.m, svrDecode.m, svrEncode.m, svrExcitationStructures.m, svrRearrangeAxes.m, svrSetUp.m, svrSliceWeights.m, svrSolveDPack.m, svrSolveDVolu.m, svrSolveTVolu.m.

./SVR/Common

contains common functions used by SVR methods: computeDeformableTransforms.m, finalizeConvergenceControl.m, initializeConvergenceControl.m, initializeDEstimation.m, modulateGradient.m, prepareLineSearch.m, updateRule.m.

./Alignment

contains functions for registration.

./Alignment/Elastic

contains functions for elastic registration: adAdjointOperator.m, adDualOperator.m, buildDifferentialOperator.m, buildGradientOperator.m, buildMapSpace.m, computeGradientHessianElastic.m, computeJacobian.m, computeRiemannianMetric.m, deformationGradientTensor.m, deformationGradientTensorSpace.m, elasticTransform.m, geodesicShooting.m, integrateReducedAdjointJacobi.m, integrateVelocityFields.m, invertElasticTransform.m, mapSpace.m, precomputeFactorsElasticTransform.m.

./Alignment/Metrics

contains functions for metrics used in registration: computeMetricDerivativeHessianRigid.m, metricFiltering.m, metricMasking.m, msdMetric.m.

./Alignment/Rigid

contains functions for rigid registration: convertRotation.m, factorizeHomogeneousMatrix.m, generatePrincipalAxesRotations.m, generateTransformGrids.m, jacobianQuaternionEuler.m, jacobianShearQuaternion.m, mapVolume.m, modifyGeometryROI.m, precomputeFactorsSincRigidTransformQuick.m, quaternionToShear.m, restrictTransform.m, rotationDistance.m, shearQuaternion.m, sincRigidTransformGradientQuick.m, sincRigidTransformQuick.m.

./Build

contains functions that replace, extend or adapt some MATLAB built-in functions: aplGPU.m, det2x2m.m, det3x3m.m, diagm.m, dynInd.m, eigm.m, eultorotm.m, gridv.m, ind2subV.m, indDim.m, matfun.m, multDimMax.m, multDimMin.m, multDimSum.m, numDims.m, parUnaFun.m, quattoeul.m, resPop.m, resSub.m, rotmtoquat.m, sub2indV.m, svdm.m.

./Control

contains functions to control the implementation and parameters of the algorithm: channelsDeepDecoder.m, parametersDeepDecoder.m, svrAlgorithm.m, useGPU.m.

./Methods

contains functions that implement generic methods for reconstruction: build1DCTM.m, build1DFTM.m, buildFilter.m, buildStandardDCTM.m, buildStandardDFTM.m, computeROI.m, extractROI.m, fctGPU.m, fftGPU.m, filtering.m, fold.m, generateGrid.m, ifctGPU.m, ifftGPU.m, ifold.m, mirroring.m, resampling.m.

./Python/deepfetal/deepfetal

contains python methods.

./Python/deepfetal/deepfetal/arch

contains python methods to build deep architectures: deepdecoder.py.

./Python/deepfetal/deepfetal/build

contains python methods with generic functions: bmul.py, complex.py, dynind.py, matcharrays.py, shift.py.

./Python/deepfetal/deepfetal/lay

contains python methods to build deep layers: encode.py, resample.py, sinc.py, sine.py, swish.py, tanh.py.

./Python/deepfetal/deepfetal/meth

contains python methods with generic deep methodologies: apl.py, resampling.py, tmtx.py, t.py.

./Python/deepfetal/deepfetal/opt

contains python methods for optimization: cost.py, fit.py.

./Python/deepfetal/deepfetal/unit

contains python methods to build deep units: atac.py decoder.py.

./Tools

contains auxiliary tools: findString.m, removeExtension.m, writenii.m.

./Tools/NIfTI_20140122

from https://uk.mathworks.com/matlabcentral/fileexchange/8797-tools-for-nifti-and-analyze-image

NOTE 1: Example data provided in the dataset svr_inp_034.mat. For runs without changing the paths, it should be placed in folder

../RGDSVR-Data

Data generated when running the example script appears in this folder with names svr_out_034.mat and x_034.mat.

NOTE 2: Instructions for linking the python code in loadPythonDeepFetal.m.

NOTE 3: pathAnaconda variable in rgdsvr_example.m needs to point to parent of python environment.

NOTE 4: Example reconstruction takes about half an hour in a system equipped with a GPU NVIDIA GeForce RTX 3090.