WO2016042037A1 - Bias correction in images - Google Patents

Bias correction in images Download PDF

Info

Publication number
WO2016042037A1
WO2016042037A1 PCT/EP2015/071248 EP2015071248W WO2016042037A1 WO 2016042037 A1 WO2016042037 A1 WO 2016042037A1 EP 2015071248 W EP2015071248 W EP 2015071248W WO 2016042037 A1 WO2016042037 A1 WO 2016042037A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
bias
registration
images
bias correction
Prior art date
Application number
PCT/EP2015/071248
Other languages
French (fr)
Inventor
Bin Zou
Akshay Sadananda Uppinakudru PAI
Lauge SØRENSEN
Mads Nielsen
Original Assignee
Biomediq A/S
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority to GB1416416.4 priority Critical
Priority to GBGB1416416.4A priority patent/GB201416416D0/en
Application filed by Biomediq A/S filed Critical Biomediq A/S
Publication of WO2016042037A1 publication Critical patent/WO2016042037A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00Image enhancement or restoration
    • G06T5/007Dynamic range modification
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00Image enhancement or restoration
    • G06T5/50Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/30Determination of transform parameters for the alignment of images, i.e. image registration
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10072Tomographic images
    • G06T2207/10088Magnetic resonance imaging [MRI]
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30016Brain

Abstract

Computer analysis of data sets representing images such as MRI images to achieve bias correction and image registration, each image including a bias in intensity within the image of unknown magnitude, is performed by: a) inputting a digital data set of a first image and a digital data set of a second image into a computer; b) calculating a deformation of said first image that transforms said first image into a transformed image that is an optimised approximation of said second image and c) simultaneously calculating and applying a bias correction which is applied to said first image and a bias correction which is applied to said transformed image such that each of the first image and the transformed image is individually corrected for bias therein. Generally, an average of the bias correction over the first image is equal and opposite to an average of the bias correction over said transformed image.

Description

Bias Correction in images

The present invention relates to methods for the computer analysis of data sets representing images in order to correct for intensity bias of unknown magnitude within each image. The invention has particular, but not exclusive, relevance to processing MRI images.

Structural magnetic resonance imaging (MRI) is a noninvasive technique for examining the physical structure of the brain (for example, calculation of the volumes of tissue) . This is of high value for monitoring neurodegenerative diseases such as dementia where it is well known that the brain shrinks as the disease evolves.

A common approach to measuring volume change is by use of image registration (a method that aligns two images by deforming one of the images) . By analyzing how the image has to be deformed to match the other image, change in volume can be computed. In order to best analyze MRI scans, they need to be free of any artifacts. MRI scans are usually corrupted with noise from various factors arising from the scanner. These normally appear as smoothly varying intensities in the images, so called bias fields, which do not represent the human anatomy. Such bias fields can severely distort the analysis.

Computational anatomy tools such as image registration have been ubiquitous in characterizing the longitudinal variations in the human brain. One such application is utilizing the atrophy scores obtained from an image

registration to infer statistical differences in different diagnostic groups such as Alzheimer' s disease (AD) and normal controls (NC) . In this context, tensor-based morphometry

(TBM) has been widely used in clinical trials to estimate atrophy by analyzing the Jacobian determinant maps constructed from the obtained deformations [1] . However artifacts in images may alter atrophy measured from image registration and can subsequently hamper the power of a clinical trial [2] . A typical artifact seen in brain magnetic resonance imaging (MRI) scans is a smoothly and slowly changing spatial

variation in image intensity, or bias, which is caused by factors such as radio frequency excitation field, magnetic field inhomogeneity, and non-uniform reception coil

sensitivity [3] .

Image uniformity can be improved by non-differential or differential bias correction methods. Among non-differential bias correction methods, the initial ones followed a

prospective approach where bias was assumed to be systematic image acquisition errors [4] . An alternative and effective approach was to correct images after the acquisition using standard image processing tools [4] . The most popular method among the latter is nonparametric non-uniform intensity

normalization (N3) method, which modeled bias as a smooth multiplicative field and acquired the bias by maximizing high frequency content of the image histogram [5] .

Differential bias correction was first introduced in [2] . Based on the assumption that bias should have large-scale structure in the difference image of rigidly registered

longitudinal MRI scans, unsharped bias was removed using median filter. A method extended this from two images to multiple images by using the geometric means of the pairwise differential bias [6] . However, because bias was estimated from scans after rigid registration, the methods also removed some true intensity variation, since differential bias from scans also included intensity difference due to atrophic tissue shifting [2] . It is likely that differential bias correction methods after nonrigid registration would perform better. Some works formulated the global transformation by multiplying a bias term to the mean of the serially aligned images [7], while others formulated the local transformation by multiplying the bias term to one image, either source image or target image [8-11] .

However, the bias field affects both source and target images when recorded. If only bias in one image is corrected, the result of a (non-rigid) registration will depend on such common multiplicative field and thereby be affected by whether one or two images are bias corrected. In practice, this may be seen e.g. in the cerebellum - it often appears that this region of the brain is not correctly bias corrected in one of the images and opposite atrophy patterns in white matter and grey matter may appear [12] . The accuracy of the atrophy scores in presence on bias therefore becomes largely dependent on the choice of image on which the bias field is modeled.

There is a continuing need to develop better methods of correction of bias in images to allow better automated

measurement of differences between the images that are not due to bias.

The present invention now provides a method of computer analysis of data sets representing images to achieve bias correction and image registration, each image including a bias in intensity within the image of unknown magnitude, the method comprising :

inputting a digital data set of a first image and a digital data set of a second image into a computer;

in said computer calculating a deformation of said first image that transforms said first image into a transformed image that is an optimised approximation of said second image and

simultaneously calculating and applying a bias correction which is applied to said first image and a bias correction which is applied to said transformed image such that each of the first image and the transformed image is individually corrected for bias therein.

Preferably, an average of the bias correction over the first image is egual and opposite to an average of the bias correction over said transformed image either exactly or within a percentile sufficiently small to ensure that the sum of the averages is represented within the computers precision of floating point numbers . Each said data set may represent an MRI image.

Alternatively, each could relate to an image obtained by Computed Tomography, e.g. by X-ray, or by Nuclear Imaging such as PET or SPECT, microscopy images of e.g. histological slices.

Said MRI image may be of a brain or a part thereof.

Each said data set may represent an image of the same object from the same viewpoint obtained at spaced time points, for instance each may be an MRI of the brain of the same patient separated in time, e.g. by a year or more.

Alternatively, each said data set may represent an image of a different instance of the same type of object from the same viewpoint. For instance, the images may be a set of brain MRI images from different patients.

One said image may then adopted as fiducial and each of a pair of other said images is subjected to bias correction and registration with said fiducial image, whereby to better define the bias correction for said fiducial image.

Alternatively, multiple images may be bias corrected and registered to produce a new image which is a bias corrected centroid of the multiple images. This may be done by iterating an update of the centroid image applying first the average

deformation obtained by registrations to the multiple images and secondly applying the average of the bias field obtained while registering to the multiple images. This process is continued until convergence is obtained as both the average registrations and average bias fields are negligible.

Optionally, simultaneous bias correction and said deformation calculation is conducted by optimising an objective function which comprises a term which is minimised when the images are optimally registered. This may be for instance Mutual Information or

Normalised Mutual Information or cross correlation. Preferably, said objective function further comprises a regularisation term which penalises deviation from linearity of calculated parameterised bias and deformation fields.

Preferably, said objective function further comprises a term constraining the average bias applied to each image to be

essentially or exactly equal and opposite, as described above.

In a further aspect, said invention comprises a computer programmed to accept as input data sets representing images and to process said data sets to achieve bias correction and image registration, each image including a bias in intensity within the image of unknown magnitude, said processing in said computer comprising calculating a deformation of said first image that transforms said first image into a transformed image that is an optimised approximation of said second image and simultaneously calculating and applying a bias correction which is applied to said first image and a bias correction which is applied to said transformed image such that each of the first image and the transformed image is individually corrected for bias therein.

From the bias corrected and registered images an estimate of change of size of said object between said time points may be obtained. Where the images are time separated brain images, such a measure of atrophy may be of diagnostic significance in

detecting Alzheimer's disease or in forecasting its onset.

Differential bias correction is an important tool when simultaneous assessments of longitudinal scans are made.

Among the existing methods, the bias model is applied to only one of the images. This may lead to inconsistent atrophy estimation depending on which image it is applied to. In preferred embodiments of this invention, we use a B- spline free-form deformation based two-image differential bias correction method where both images in the

registration process are corrected for bias simultaneously. Further, symmetry in bias correction is achieved via a new regularization term. In a simulated experiment, described below reproducibility of atrophy measurements m a single- image bias correction method largely depended on the choice of the image that was corrected while this choice did not matter with the two-image bias correction method of the invention. On Alzheimer's disease neuroimaging initiative data, the two-image bias correction method performed b when compared to registration of separately bias corrected images .

The invention will be further described and illustrated by the following description of preferred embodiments with

reference to the accompanying drawings in which:

Figure 1 is a graph that shows results obtained on whole brain images in an evaluation of the invention described below.

Figure 2 is a graph that shows results obtained on

hippocampus images in an evaluation of the invention described below .

Figure 3 shows a graphical representation of a 1-D Wendland kernel of different scales and orders.

Figure 4 shows Kernel Bundle Framework for SVFs .

In image registration, the difference between the images due to bias fields can be accounted for as an integral part of the alignment process. Existing registration tools do this by trying to account for the bias field in one of the images and evolving the other image to the same intensity level as the uncorrected image .

However, this approach is asymmetric, because it matters which image is selected as the one where bias is accounted for. Moreover, the real bias field is not necessarily removed, only the intensities are changed to match in the two images. It is therefore possible to add bias to the uncorrected image instead of removing bias from the evolving image .

The present invention aims at correcting both images simultaneously during the registration procedure. The proposed method involves assuming two bias fields, one in each image. During the registration optimization, the images are aligned after accounting for the above-mentioned bias fields. In order to be consistent with both the images it is preferred to enforce that an at least essentially equal amount of bias correction is done in both images. This criterion may be adequately met if the correction applied to one image is of same order of

magnitude as the correction applied to the transformed image, so that the sum of their averages is well represented within the computers precision of floating point numbers.

In a preferred methodology according to the invention, applied to measuring atrophy in brain MRI images separated in time, in order to remove a dependency of atrophy scores on which image is bias corrected, we propose a B-spline free-form

deformation based two- image differential bias correction method where both images in a registration process are corrected for bias simultaneously. Both the images are treated with

independent bias models; however, the models are consistently applied through a regularization term. The method may be directly applied to any gradient descent based non-rigid

registration method.

In the following, Scalars are normal typeface. Vectors are boldface. Spatial coordinates are given by x = ,γ

Given two images, image registration tries to find a transformation T that best maps point in the floating image to the corresponding point in the reference image. In this embodiment, the transformation is a combination of a global transformation and a local transformation. The global transformation model is a rigid registration using 6 degrees of freedom. Following the rigid registration is a refinement using non-rigid registration where the transformations are modeled using free-form deformation framework (FFD) [13] . FFD consists of three components: a deformation model, an objective function and an optimization scheme. Combined Deformations and Bias Model

In geometric deformation model, a cubic B-spline interpolation scheme is used as in [13] . We manipulate a mesh of geometric control points pljh overlaid on the image, with spacing δ , , and ^respectively along the x-, y-, and z-axis, to represent the image.

Bias fields are modeled on image intensity by

multiplying a scalar exponential term to both the reference image and deformed floating image. Here the exponential terms are used, since they are positive and the derivatives are simple and cheap to compute. B-spline basic function is used for image intensity interpolation.

To define the bias model, we denote the domain of the 3D brain MRI scan as

V = { (x, y, z) I 0 < x < X, 0 < y < Y, 0 < z < Z} .

The multiplied bias fields to the reference image R and deformed floating image F (T) are defined as eBR and EBF (T) _ Let hlmn and bdef be the mesh of control points overlaid on R and F (T) , with spacing δx bR , δy bR , <5*s and δb , yb ' <5*F respectively, then BR and BF (T) field for each voxel (x, y, z) can be computed as

Figure imgf000009_0001

flF{T(x)) =∑ /33 - d)β* (3 - e) 33 - /) (2) where 3J is the cubic B-splme basis function, I, m, n and d, e, f are control point indexes along x-, y-, and z-axes respectively for two images in V and T (x) =

(Tx(x), Ty(x), Tz(x)) .

Equation (1) gives the value of the bias field in the position x (real valued, not necessarily integer valued) . It is given as a super position of 3-dimensional 3RD order b-spline basis functions computed as the product of one-dimensial b-splines (β3(χ) times β3 (y) times β3(ζ)) centered in the coordinates l,m,n in the three dimensions (χ,γ,ζ) =x, weighted by the scalar jbimn . The functions β3 is a convenient choice of functions (third order b-splines) from the literature. bimn are parameters. The bias field is used to create a bias-corrected image

Rcorrected (X) = R(x) exp ( BR(X) )

The problem is to determine the bias field: determining the parameters . imn (and the parameters i¾ef and p±j k introduced later) .

Equation (2) is identical to (1) except that it parameterizes the bias field of the floating image which is moved from x to (x) according to a transformation T: R3 →R3 that maps three-dimensional space to three-dimensional space. Now, a corrected floating image is

Fcorrected(T (X) ) = F(T(x)) exp ( BF(T(x)))

If this holds it trivially follows that also the following holds

Fcorrected (X) = F(x) exp ( BF(x))

Hence, by using BF ( (x) ) , and applying this to the transformed floating image F ( (x) ) , one implicitly applies the bias field exp ( BF (x) ) to the untransformed floating image . Hence the bias field acts in a coordinate system attached to the scanner

(untransformed) and not attached to the patient (transformed) . This is physically more realistic, as bias fields originate more from scanner artefacts than patient artefacts.

Objective Function

Deformation model parameter ±jk bias field parameters for reference image blmn and deformed image bdef are used as parameters to optimize the objective function (3) which is a combination of normalized mutual information (NMI) term, regularization term and symmetry term. The NMI term is used as the similarity criterion to measure the alignment between two images, the regularization term is used to smooth the deformation, and symmetry term is used to achieve symmetry .

C = -NMI + R + 7S (3) Equation (3) is the overall penalty function to minimize in order to solve the problem (i.e., estimating warp and bias fields simultaneously) . It consists of three terms weighted against each other. C is a function of all

deformation field parameters pijk and all bias field

parameters j def and j imn . NMI and R are dependent on all

parameters whereas S depends on the bias field parameters j def and j imn only. The parameters obtaining the minimal

value of C, would not change if C had been multiplied by a constant. Hence, a weighting parameter on NMI has been omitted as only the relative weighting of NMI, R, and S is of importance.

The regularization term R is based on the discrete

Laplacian in the geometric deformation controls points and bias field control points [14] and is given by

Figure imgf000011_0001

Equation (4) is the regularization term of (3) . It is the sum of the regularization of the three b-spline parameterized deformation and bias fields. It is implemented as terms trying to keep the fields as close as possible to linear functions by penalizing differences to a linear function. The operator "Δ" denotes the discrete laplacian so that

Δ -folmn = 1/6 ( 6* j imn ~~ -fol-1 mn _ -fol+1 mn _ -blm-1 n _ -folm+1 n -folmn-l -folmn+l )

This is j imn from which is subtracted the average of its 6 neighbours in 3 dimensions. Making this close to zero (by penalizing its square) in all points enforces the b values to be close to a linear function of its indices . The special linear case ( j imn = a*l+ b*m +c*n+d) has a zero discrete laplacian ( Δ bimn = 0) . Hence, the j imn cannot vary wildly, but must stay relatively close to a function like a*l+ b*m +c*n+d. The first term of (4) could more stringently have been written using

Apljk = (Apx ljk , Apy ljk ,Apz ljk )

using the fact that p = (px, py, pz) and

I Apljk 12 = Apx ljk Apxi j k + Apy ljkApy ljk, + Apz ljkApz ljk showing that the regularization of the transformation T ( x ) is sought to be close to a linear function in the three coordinate functions

independently .

Bias field intensity value could go to zero or

infinity, to avoid this the symmetry term is proposed

Figure imgf000012_0001

Equation ( 5 ) expresses the preference for the average log- bias-field in reference and floating image to sum up to zero.

That is, they should on average be equally large with opposite signs although locally within the registered images the sum of the bias fields may not be zero. This can be obtained by adding a constant to both all binm and £>def . This does not influence R and NMI as they are invariant to such constants multiplications of the images. Hence, this term S will in the minimum always be

identical to zero and the symmetry exactly fulfilled. The purpose of ensuring that the two bias fields are equally large with opposite signs, is to ensure that, we do not, in the minimum identified by gradient descent, add a very large positive or negative constant to both log-bias-fields. If we added or subtracted such constants, we maybe put some numerical imprecision into the system.

The NMI term which measures the amount of

information that two images share using entropies [ 1 5 ] , is formulated as

I(e4x¾eBFiT)xf(T)) l J where the numerator of (6) is the sum of marginal entropies of two images, and the denominator of (6) is the joint entropy of two images.

Equation (6) expresses the "Normalized Mutual

Information" (NMI) . This is a variant of the Mutual

Information MI. Assume two random variables X, Y, then the MI(X,Y) = H(X) + H(Y) - H(X,Y), where H() is the entropy of a random variable. The more uniform a distribution is, the larger is the entropy. For image registration we wish to maximize the mutual information of the intensity in the two images taken in corresponding positions. That is, assuming that H (X) and H(Y), the entropy of the intensity distribution in the two images, does not vary with how the images are registered, the Mi-based registration will minimize H(X,Y), that is to make the joint distribution of the corresponding intensities as non-uniform as possible. That is, it wants the joint histogram of the two images to be placing the intensities in as few bins as possible.

This is assumed to be obtained if corresponding points are images of the same physical point. From a source coding perspective, the MI tells how many bits of information X tells about Y (or vice versa) . That is: if one wishes to encode an intensity from one image, how many bits may one on average save, if one already knows the corresponding intensity in the other image.

The Normalized Mutual Information is defined as

NMI = (H (x) +H (Y) ) /H (X, Y) .

This has been published as a way to handle the problem that images are of finite extent, and thereby there may be a tendency using MI to have only few overlapping points in the images. In this situation MI can, due to the low number of samples, become artificially high. NMI may be seen as a hack without root in

information theory, but it is applied in practice with great success.

For the purpose of the invention it is not expected that it makes a major difference if MI or NMI is used.

Indeed, instead of NMI, any data term that is obtains its minimal value when images are optimally registered may be used. MI and NMI are especially preferred as they are not just maximizing H(X,Y), but also simultaneously trying to minimize H (X) and H (Y) .

Minimizing the histogram entropy is another well- known methodology to perform bias correction of a single image .

Also, instead of NMI, the data term may be the cross- correlation :

CC(R,F) = J R(x)F( T(x)) dx / [ sqrt ( JR2 (x) dx)

sqrt ( JF2 ( T (x) ) dx) ]

In order to compute the marginal and joint entropy in (6), an approach using parzen windows [16] is used. Denote the index of the histogram bins as κ

and L, the joint histogram is

, , . _f31 eD"{x) x R(x) -R°\

x€ ^ R '

Figure imgf000014_0001

here a is the normalization factor, R° and F° are the minimum intensity value of two images, and AbR and AbF are the intensity range of each bin. The corresponding marginal histograms can be obtained by integrating (7) .

Equation (7) shows how in this embodiment in practice is obtained the joint histogram (here denoted p, n.b. this has nothing to do with the transformation control points pijk, but is used as it is assumed to have something to do with a joint distribution of intensities) of two images with correspondence defined by p±jk and bias correction defined by .bimn and i¾ef. After bias-correction, image intensities are not any longer integer values, and a discrete histogram is not expedient to use due to one effect: there is truncation error when an intensity is put into bins. This makes the MI as a function of bias-correction a piecewise constant function separated by jumps in value when the change of bias correction make a pixel jump to a neighboring bin in the histogram. This makes gradient-based optimization impossible. Instead, a b-spline, β3 is centered in each floating point intensity and represents the joint histogram (This b-spline is the same mathematical function, but should not be confused with the basis-functions used for the bias fields or the deformation field. ) . This has the advantage that the histogram p becomes a differentiable function in the parameters pijk, bimn and i¾ef. Thereby also NMI (or MI) becomes differentiable in the parameters, and a gradient-based optimization methodology may be used.

Optimization Scheme

The limited memory Broyden-Fletcher-Goldfarb-Shanno (L- BFGS) algorithm [17] is used to optimize the objective

function. In order to proceed the optimization, we need to compute the derivative of the objective function (3) with respect to Pijki blmn and bdef. To proceed optimization, the derivatives of joint and marginal entropies and histograms have to be evaluated. The derivative of (7) with respect to Pijk can be computed as :

Figure imgf000016_0001

(8)

Figure imgf000016_0002

The derivative of (7) with respect to j ijm can be computed as :

Figure imgf000016_0003

(3)

Figure imgf000016_0004

The derivative of (7) with respect to bdef can be computed as :

dbdef

Figure imgf000016_0005

The choice of adding the bias correction term before or after transformation in the floating image is a matter of choice. We chose to multiply the correction term after the transformation in order not to interfere with the spatial deformation model. However, this choice is not expected to have a significant influence on the final registration. Equations (8,9,10) give the derivative of (7) with respect to the parameters p±jk, j imn and . def respectively. The derivatives of (3) are assumed to be easy to find when these are given. dC/dq (where q here could be any parameter) is

dC/dq = 3NMI/3q + Λ dR/dq + γ dS/dq

3NMI/3q = [ (3H (X) /dq+dH (Y) / 3q) H (X,Y) -3H (X,Y) / 3q (H (x) +H (Y) ) ] /

H (X, Y) 2

and as H (X) =- fp (x) log p (x) dx we find after some algebra

3H(X)/3q = -Jdp (x) /3q (log p(x)+l)dx

Similar equations hold for H (Y) and H(X,Y) and the

derivatives of R and S are quite trivial to find as they are simple sums of squares or squares of sums.

Parameters

Three levels and five levels of transformations were used respectively in rigid and non-rigid registration.

Levels and evaluation point settings were similar to [14] . Different Gaussian blurrings were applied to the images at different levels of registration (kernel sizes of 15, 8, 2 mm for the three levels of rigid registration and 2.0, 1.5, 0.5, 0.2, 0.2 mm for the five levels of non-rigid registration) . The regularization parameter Λ and symmetry term parameter γ were empirically chosen as 0.03.

Cube Propagation

In order to evaluate the performance of registration, cerebral atrophy in serial MRI scans were estimated. Since the transformations were a composition of B- splines, cube propagation [14] was used to compute local volume changes.

Usually, given a deformation field and an anatomical mask, regional atrophy is estimated by summing the Jacobian determinant over the region of interest (ROI) . However, for registration schemes where the analytical expression of the transformation is not available, the Jacobean determinant needs to be approximated using finite differencing schemes [14a]. We instead utilize Cube Propagation (CP) to measure atrophy. Here, each face of a cubic voxel is triangulated and the volume under each triangle after transformation is summed to get the volume of the transformed cube. Tetrahedral meshing is also similar in terms of both numerical precision and meshing, however with CP one needs to only triangulate the surface which is simpler in terms of bookkeeping of the indices and computations. A detailed description of CP can be found in [31a] . Thus, cube propagation is used here to avoid numerical noise in the deformation being amplified by using Jacobian integration.

Alternative Registration methods

Registration methods other than that described above may be used whilst applying the simultaneous differential bias correction described herein in the same way. For instance, instead of utilising B-splines, one may use parameterization of stationary velocity fields for diffeomorphic registration using a class of multi-scale, multi-shape regularizing kernels called the Wendland kernels. This registration scheme is here termed wKB-SVF. We propose a framework that incorporates the best characteristics of state-of-art registration schemes: 1) we restrict the space of velocity fields to a specific class of function spaces,

reproducing kernel Hilbert spaces (RKHS) [la], [2a], [3a], [4a], [5a], [6a]; 2) we parameterize the velocity fields using compactly supported reproducing kernels inherited from the RKHS structure, and we subsequently represent the high dimensional ODE via a smaller set of control points and vectors [7a], [8a]; 3) we provide a multi-scale representation of the velocity fields using the kernel bundle framework [9a] .

In the absence of validated models for inter-subj ect/intra- subject anatomical variability, deformations characterizing anatomical changes such as change in organ growth are generally assumed to be smooth and invertible. Three popular choices of diffeomorphic deformation models are: a) Large deformation diffeomorphic metric mapping (LDDMM) , b) freeform deformations and c) stationary velocity fields (SVFs) . Among them, a) and c) naturally generate diffeomorphisms and b) requires explicit regularization terms to ensure diffeomorphic transformations. For a discussion about commonly used constraints on deformation models see [10a] .

Each of these methods involves finding an optimal

diffeomorphism that connects two images. SVFs are less

computationally expensive compared to LDDMM due to the constant velocity field assumption. Here we use SVFs together with some key concepts from LDDMM because SVFs satisfy the dual goal of

generating diffeomorphisms while keeping computational complexity low. A key feature of LDDMM is that the velocity fields are modelled on a Hilbert space. This space can be constructed using reproducing kernels, and this approach allows optimal solutions to specific optimization problems to be found as linear combinations of the reproducing kernels. We model the stationary velocity fields on a Hilbert space constructed using a class of reproducing kernels called Wendland kernels. A key property of Wendland kernels is that they are of compact support. The construction reduces computational complexity because both the deformation field and the regularization term evaluate to zero outside the support of the kernel [6a] . Existing parametric versions of SVFs [7a], [8a] use kernels where the evaluation of the energy term often requires spatial discretization (bending energy for

instance) ; Wendland kernels require no such spatial

regularization. In addition, we will use these kernels in a kernel bundle framework to provide a multi-scale modelling of the deformation field. Using the multi-scale feature from the kernel bundle framework allows us to express the combination of various scales as a simple sum while still remaining in the Hilbert space.

Figure 3 illustrates the orders of kernels that through the bundle construction can be used simultaneously in wKB-SVF. In addition, the proposed framework allows the flexibility to simultaneously optimize for scales like in the LDDMM based kernel bundle framework [9a] .

In terms of performance, wKB-SVF provides better overlaps than the registration methods considered in the study [11a] on the publicly available MGH10 dataset. On the CUMC12 dataset, this method provides significantly better overlaps than most of the registration methods except for SPMDARTEL where the difference is insignificant. The framework also separates diagnostic groups of Alzheimer's disease (AD) and normal controls (NC) better than the Freesurfer longitudinal segmentations when used to compute longitudinal atrophy scores. The results illustrate that the wKB- SVF is well suited for both inter- and intra-subj ect registration.

In flow-based registration schemes, deformations are

generated by integrating a smooth velocity field over time. Two prominent flow-based image registration frameworks are the LDDMM (time-varying ODEs) [2a] and SVF (time-constant ODEs) [12a],

[13a]. In LDDMM, the deformations can be parametrized by initial velocity fields (or their dual, momenta) and the resulting diffeomorphism paths are endpoints of the corresponding Riemannian geodesies. This particular setting is computationally expensive since it involves solving a geodesic equation on an infinite dimensional group. An alternative to LDDMM are the SVFs . Here, the diffeomorphisms are one parameter subgroups parametrized by time-constant velocity fields through the Lie group exponential. The Lie group exponential is realized as a time-integration of the velocity field. The time integration is usually approximated using integration schemes such as Euler' s or scaling-and-squaring [12a]. The generated diffeomorphism paths are geodesies with respect to the canonical Cartan connections [14a]. The main drawback of SVFs is the lack of metric on space of diffeomorphisms which is important for performing statistics such as PCA [15a] or regression [16a]. SVFs were initially proposed by [12a] and were further utilized with modifications in [17a], [7a], [8a], [13a], [14a]. Among these, [17a], [13a], [14a], [18a] use the entire image space for dissimilarity minimization. In studies [7a], [8a], the velocity fields are instead parameterized by interpolating kernels like b-splines. A thorough overview on existing registration schemes can be found in the study [10a], [11a]. Here we will restrict our focus to flow-based registration schemes and specifically to the parameterization of the stationary velocity fields (SVF) .

In both SVF and LDDMM, the vector fields belong to a subspace V of square integrable functions, ,£2. The subspace V is generally completed using a Hilbert norm induced by a differential operator [19a] . With sufficient conditions on the operator, the space is a reproducing kernel Hilbert space (RKHS) [20a] : the Riesz

representation theorem states that every linear form arises as an inner product with the representer. The representer is the reproducing kernel. A linear form is an evaluational functional that provides a mapping of a vector to It. In case of finite dimensional optimization problems, RKHSs allow evaluation of the optimal space in terms of the reproducing kernel itself: for instance, for an interpolation problem defined as to find E V of minimum norm that satisfies v( r,) ¾ βι> β· <= K, the solution can be expressed as the regularized optimization problem (Ref. theorem 9.7 [20a]), I v\ \ \ + C∑ |v(jr,) - £|2. The solution then attains the form, v =∑^ A"(jr, i . Here K is the reproducing kernel. The whole problem is thus reformulated to a finite dimensional optimization problem involving only the vectors .

In contrast to the most common approach, we will take advantage in the fact that there is flexibility in choosing reproducing kernels directly as opposed to being imposed by an operator. This approach allows us to minimize computation through the use of compactly supported kernels. Alternate options to parameterizing velocity fields are by either using B-Splines or truncated Gaussians. The latter is no longer continuous and the evaluation of the energy term of the former (like bending energy) is an approximation since it depends on spatial discretization. Wendland kernels [21a] (the choice of reproducing kernels used here) on the other hand emulate B-splines in both computational complexity and smoothness. In addition, they also provide the necessary mathematical properties (smooth, Ck for some k, norm- minimizing) to realize a diffeomorphic transformation model. The role of reproducing kernels and the corresponding regularization in the context of LDDMM has been explored in [9a], [6a], [3a] .

Inter-subject registration often requires smooth yet large deformations, whereas intra-subject registration requires deformations at much smaller scales. For example, anatomical changes in the hippocampus are often minute and changes in ventricles (cerebrospinal fluid) may require large deformations. It is desired that such deformations are recovered using the same transformation model. The type of the resulting deformation is restricted by the scale of the parameterizing kernel. Scale, in this case, can be interpreted as either the support of the kernel or the spacing between the control points. If the scale of the kernel is large, then matching of the larger structure may be goo and the transformations smooth. However, the matching of smaller structures like the hippocampus may not be satisfactory. On the other hand, if the scale is small, the matching may be good but the resulting transformation is spiky and may lead to undesirably large Jacobians [3a] . One way to handle such variability in deformation scales (also to avoid local minima' s in optimization) is via a pyramidal approach i.e., by changing the scales of image smoothing or the resolution of the control points. This approach however, is still limited by the range of deformations achievable by the shape and size of the kernel. The kernel bundle framework handles this by providing a scale-space representation of the kernels. A very attractive feature of the kernel bundle framewor is that the representation of the multi-scale kernel is a simple linear combination of kernels of different support or resolution. Standard parameterizations of velocity fields like ones using B- Splines require additional routines such as knot splitting to combine various scales of the velocity field. The idea of exploiting RKHS kernels to build multiscale kernel based

diffeomorphic image registration is not new in the context of LDDMM [9a], [3a] . However, the application of kernel bundle framework in the context of SVF based image registration is believed to be novel. Further, a combination of compactly supported reproducing kernels and the kernel bundle framework has not been explored in diffeomorphic image registration.

We start by describing SVF based image registration and presenting the application of RKHSs in the context of SVFs. We then discuss how computational complexity can be minimized by representing velocity fields with compactly supported reproducing kernels. Followed by this, we discuss the adaptation of the kernel bundle framework to SVFs together with compactly supported Wendland kernels.

Given a floating image I and a reference image l2 with a spatial domain Ω € d , image registration involves finding a transformation ψ : £1 X R → Ω that aligns the images. The transformation is found by minimizing a dissimilarity measure between the images under certain constraints encoded in a regularization term. A general cost function is of the form: arg vcan{E(llrI^)) = arg muiED (l^tp), J2) + l_¾ ( ?) + ICCEICC( ) (la ψ ψ where A, [CC are user-specified constants controlling the degree o regularization, ED is a dissimilarity measure that allows comparison of the floating image to the reference image, ER is a regularization term that encodes desired properties of φ, and EiCC can be included as an additional penalty term to enforce inverse consistency, see discussion of Equation 6a. The regularization term can either be explicitly minimized as in the parametric approach or can be implicitly restricted by convolving with a low pass filter [13a]. The transformation now is restricted to the group of diffeomorphisms Diff (Ω) . In flow-based schemes, a time- dependent velocity field v(jr,£) : Ω X R→ Rd is integrated to obtain a displacement. The governing differential equation is of the form

3φ(χ, t)

dt where φ is the displacement and JQ | £)| j ljB0 dt < co where [Q,T] is the time interval. The path of diffeomorphisms f?(-,t) is in practice obtained by numerical integration. Solving the non-time stationary differential equation is generally computationally expensive .

With stationary velocity fields (SVF) [14a], the velocity field v(¾,t) is constant in time. The paths parameterized by SVFs are exactly one parameter subgroups of Diff (Ω) . These paths are quite different from the Riemannian geodesies in the sense that the paths are metric-free [14a] . Let £1 be the spatial domain of JL with x E Ω as a spatial location. Let G c Diff(Q) be a subspace containing the diffeomorphic transformations, and let V be the tangent space of G at identity Id containing the velocity fields v. A path of diffeomorphisms is generated by the stationary flow equation,

dq>(x,t) (2a) — = v(p(x, t}) with init ia1 condition <p(Xj 0) = x. The final transforrriation φ(χ) = φ(χ, ) is the Lie group exponential map Exp (v) . This Lie group exponential can be approximated by Euler integration [7a]. For example, given p steps and φ — φ(χ, t), the Euler intecjration is ψ = x, (3a)

(4a) ψρ = x +

V

1

f. (5a) ψ w = φι

In the study [12a], the scaling-and-squaring method to exponentiate velocity fields was proposed. Here the final deformation was estimated by composing successive exponentials. However, a major drawback to this method is that at every squaring step the velocity fields need to be re-interpolated at integer positions. This may lead to undesired smoothing (interpolation) in the velocity field over which there is no apparent control.

Bossa et. al., [23a] point out the instability in the convergence properties of scaling-and-squaring. Therefore, we choose the relatively stable forward Euler's scheme for integrating the velocity fields .

In a continuous setting, diffeomorphisms generated by SVFs are invertible transformations with differentiable inverses.

However, due to the numerical integration of the velocity field, inverse consistency is not achieved in practice and needs to be explicitly enforced typically through a regularization term. In [24a], inverse consistency was enforced by penalizing the

displacement error generated after composing a transformation with its inverse. However, in this method, the computation of the inverse is a computationally expensive approximation [25a] .

Forward transformations are first computed and then the inverse transformations are approximated. We will maintain a single parameterization of the velocity field. Both the forward and backward registration are performed simultaneously. The inverse consistency term is computed as,

Figure imgf000025_0001
where φ = φ(χ, i) = Exp(v ) and φ = Exp(-v) are the backward and forward registration transformations.

In this section, we present multi-scale parametrization of velocity fields using compactly supported reproducing kernels. Similar to representations with non-reproducing kernels such as B- splines, the kernels have compact support but unlike B-splines, the reproducing property of the reproducing kernels ensure that the kernels are norm-minimizing.

In SVF based image registration, the velocity fields v, are chosen to belong to a subspace of absolutely integrable functions in ,£2. To complete this subspace, the norm associated with an appropriate differential operator L,(urv)v = (Lli,v) s, ν,,ν€ V is utilized. Usually, L is chosen to be a diffusive model of the form L = Id — aV2 [19] where V2 is a Laplacian operator. Other choices for the operator exist and discussion on them can be found in

[26a] .

The operator L provides a mapping of the velocity field v from V to its dual space V". When V is admissible [20a] , the dual space contains linear evaluational functionals §x : v→ v( ) . The evaluation functionals, that for each x E Q provide a mapping of the vector space to R, can be written as (S \v) = V{J) . According to the Riesz representation theorem, there exists spatially dependent kernels Kx = K{-,x) : Ω X a→ dxd such that v(jf) = (KxJv)v = (¾|v). This implies that (Kx,v)v = (LKx\v) and LKX = ΰχ

[20a] . Note that (·,·) denotes the inner product and ( · | · ) denotes evaluation of a functional functional i.e., (<5|t?) = i(t') where δ E V and v E V. We can therefore view K as an inverse of L . In fact, the kernel is also a Green's function with respect to the differential operator L . If the operator is differential, then K is positive definite. As a consequence, if K is constructed from a differential operator, then K is always of infinite support

[27a]. It may be computationally intensive to evaluate the deformation field and the norm if velocity fields are

parameterized using kernels of infinite support.

The approach in the previous section essentially involves first finding a mapping from V to V, and then constructing kernels that provide a mapping back to V . We will use the significant benefits in taking the reverse approach: instead of constructing kernels from differential operators which force the support of the kernels to be infinite, we choose the kernels directly. This particular arrangement allows the use of kernels to intentionally minimize computation via the compact support.

Following Moore-Aronszaj n theorem [28a]: for every symmetric positive (semi) definite kernel K exists a unique RKHS that has K as its reproducing kernel. The corresponding RKHS is the

completion of the linear space spanned by the functions of the form,

Figure imgf000027_0001
for all choices of E which is the parameter attached to each kernel centered at the points jr.6 Q. The inner product on this space provides the reproducing property such that,

(7a)

This essentially implies that we can choose an appropriate symmetric semi-positive definite kernel with compact support and this kernel has a unique RKHS associated with it. The Gaussian kernels is an example of reproducing kernels. However, due its infinite support, parameterizing velocity fields with Gaussians in dense image matching may be expensive. Using reproducing kernels to generate transformations is not new in LDDMM. Studies usually [9a], [3a], [6a] utilize reproducing kernels such as Gaussian kernels to parametrize velocity fields.

A regularization term usually is required to ensure

sufficient smoothness in the solution of the ODE. In flow-based registration schemes, this term is usually formulated as the squared norm on the velocity field. Given a reproducing kernel, the evaluation of the squared norm is simply the kernel product and does not depend on any spatial discretization like other regularization terms such as bending energy of B-Splines. The regularization term may be evaluated as,

and by linearity of the inner product and the reproducing property associated with the corresponding RKHS (7a), the norm on linear combinations of the kernel can be evaluated by

Figure imgf000028_0001

Because of this reproducing property, it is often useful to parametrize the optimal function directly using these kernels since optimal solutions are linear combinations of the reproducing kernels of the norm. In case of infinitely supported kernels like the Gaussian, the entire double sum needs to be evaluated which can be computationally expensive depending the number of kernels. In contrast, with finite support, the kernel product is zero outside the support making the evaluation of (8a) efficient. We will now outline extending the kernel bundle concept [9a] to compactly supported reproducing kernels and use it in the SVF framework .

The reasoning behind the need for a multiscale representation of a deformation has been well discussed [3a], [9a]. In brief, image deformations often occur at different scales. For instance, in inter-subject registration large scale transformations may be required and in intra-subject registration relatively smaller scaled deformations are required. For example, the deformations around hippocampus can be small while in regions like

cerebrospinal fluid, the deformation may be larger. The key is to obtain computationally efficient representations of

transformations without limiting the range and capacity of the deformation. This can be achieved by combining multiple kernels at multiple scales in the same registration framework. Typically in a kernel-based image registration scheme, the support (or scale) of the kernel is fixed. For instance in cubic-spline, the support is fixed to four by design.

The kernel bundle framework in LDDMM [9a] incorporates multiple scales of kernels as a sum in the same optimization function. We utilize the fact that the sum of multiple RKHS spaces is still RKHS. We set to achieve a similar construction with SVFs . We extend the concept of the space of velocity fields

V to a family of spaces of velocity fields V . We consider r spaces where each Vm is equipped with a norm = l...r. The

Figure imgf000029_0001

velocity fields are linear sums of individual kernels at r levels. It is represented as,

r r "m

Figure imgf000029_0002

Here v is the final velocity field, Km is the kernel at each level and a™ is the parameter associated with it. The variable im is the kernel centers at each level and TV is the number of kernels at each level. Note that Κ^χ^,χ) = A'(— —-), where qm is the support of the kernel at each level. The expression of the cost function (la) in a kernel bundle framework can be written as below :

( arg miiiv

Figure imgf000030_0001
72(Exp(vi)) + E(l2, Ii(Exp— (vi )),

arg minv E(I I2(Exp(v1 + v2)) + E(I2, (Exp - (vx + v2)),

[ arg minv

Figure imgf000030_0002
-vm))

The kernels at each level can be of any support. For instance, one can have infinitely supported Gaussian kernels in a coarser registration scale and have compactly supported kernels handle finer resolutions in the registration. Figure 4 illustrates the kernel bundle framework. Different curves represent different classes of kernels.

In this section, we will describe the compactly supported Wendland kernel [21a] used in the parameterizing the velocity fields. Note that compactly supported reproducing kernels can also be constructed instead of choosing them directly. One such example is found in [6a] .

Wendland kernels were originally developed for multidimensional, scattered grid interpolation. They are positive definite

functions with positive Fourier transforms and minimal degree polynomials on [0, 1] . They yield C23 (s is the desired degree of smoothness) smooth radial basis functions on W.d . Application of

Wendland kernels in landmark registration can be found in [29a]. They are defined as follows,

Figure imgf000030_0003
where s is the smoothness of the kernel, r is the Euclidean distance to the center of the kernel scaled by the support, g is the dimension given by |_d/2j = 2c + s + 1 and Is is the integral operator applied s times given as,
Figure imgf000031_0001

We will utilize two classes of Wendland kernels in the kernel bundle framework. They are,

(12a

Figure 3 shows both the linear (innermost) and C smooth Wendland kernels (outermost, middle) . They all have unit coefficient.

Note that Equation (10a) refers to the general family of Wendland kernels. We will however choose only particular kernels

( d = 3,s = 1) since they emulate B- splines both in terms of the smoothness properties and the shape of the kernel.

We will optimize the kernel bundle framework in a

hierarchical parallel fashion. For instance, in level one, only v1 is optimized, in level two both v2 are optimized, and so on and so forth. A limited memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) [30a] scheme is used for optimization. The optimization was done using the min-Func package (see

http: //www. di . ens . fr/ ~mschmidt/Software/minFunc. html, version 2012) . For optimization, we manually supply the derivatives. Normalized mutual information (NMI) is used as a similarity measure for both the linear transformation and non-linear transformation. We initialize the non-linear transformation with a linear

transformation with 9 degrees of freedom (DOF) . We refer to [31a] for formulations of the linear transformations that we use. We can write the cost function as (13a) . Note that the forward and backward transformations are represented as φ(χ, 1) = Exp(v) and φ'(χ, 1) = Exp(—v). The derivative of the cost function at each level can be derived as (14a) . In both (13a), (14a), the backward registration is computed in a similar way however by replacing v by —v. In (14a) , —jjj is the derivative of the transformation with respect to the nth parameter at rth level , e(l...p) is the composition number (p, total number of compositions)

Sc = — and D φ._Λ is the spatial Jacobian of the number of com ositions * x

previous composition. The flow chart of the registration

algorithm can be found in Algorithm 1. The parameter used for the registration can be found in Table I. For computational reasons, only every second voxel was used to evaluate the similarity measure. Note that for both inter-subject and intra-subj ect registration the same set of parameters are used. Note that all the r levels are optimized simultaneously depending on the level.

Algorithm 1 wKB-SVF Registration Algorithm

1: Affine registration with 12 degrees of freedom.

2: Non-rigid registration

Initialization, am = O ffl = 1. . . T

3: loop over the number of levels r

Smooth both floating and fixed image with a Gaussian of standard deviation .

Compute velocity field v (9a) . Compute displacement field φ = Exp(v).

Compute similarity measure NMI Compute the update Va E (14a) .

In order to evaluate the registration on intra-subj ect registration we measure atrophy (or volume change) in disease (Alzheimer' s disease) specific brain regions such as whole brain (WB) , Hippocampus (Hip) , Ventricles (Vent) and Medial Temporal Lobe (MTL) .

Experiments and Results (Obtained using the first described

Registration method)

Image Data

Test data in the experiments were obtained from the Alzheimer's disease neuroimaging initiative (ADNI) database and the chosen subjects were the same as in [18] was used. The images were given in 256 χ 256 χ 256 isotropic voxel cubes where voxel dimension was l x l x l mm.

Simulated Bias

In this experiment, we randomly picked a case from the available dataset and multiplied an artificial bias field to the image. Atrophy was then computed as function of the strength of the bias field. The form of the multiplicative bias field was:

Figure imgf000033_0001
where x, y, z are the spatial locations, σ is the standard deviation of the Gaussian and was chosen to be 38 and a is the strength of the bias field and was carried between 0.02 to 0.97.

Several sets of experiments were conducted using different bias correction strategies described below and results are shown in Figures 1 and 2. The figures show atrophy of whole brain (Figure 1), and of hippocampus (Figure 2) as a function of strength of the bias. y-axis represent % volume loss and x-axis represents the strength of the bias a. The six lines represent six different settings described as follows: a) bias was multiplied to the floating image and both images were corrected during registration (Reference- Double) , b) bias was multiplied to the reference image and both images were corrected during registration (Floating-Double) , c) bias was multiplied to the reference image and the same image was corrected during

registration (Reference-Reference) , d) bias was multiplied to the floating image and the same image was corrected during registration (Floating-Floating) , e) bias was multiplied to the floating image and the reference image was corrected during registration (Floating-Reference) , and finally f) bias was multiplied to the reference image and the floating image was corrected during registration (Reference-Floating) .

Runs (a) and (b) are in accordance with the invention.

After the registration, whole brain and hippocampus atrophy were computed using the generated deformation field. From Fig. 1 and Figure 2, we can observe that most variations in atrophy estimation were seen when the floating image was corrected for the bias (regardless if the bias was added to floating or the reference image) . Minor variations in atrophy scores were seen when the reference image was corrected for bias during registration and nearly no variations were seen when bias correction was performed on both the images

(regardless of which image was corrupted with bias) .

Application on ADNI data

Two-image bias correction based registration was

applied to the full dataset and atrophy in whole brain, hippocampus, ventricles and medial temporal lobe were

computed. For comparison, registration without any

differential bias correction was performed on the same data which was already pre-processed using Freesurfer' s

N3 bias correction method. To evaluate the performance of both the methods, diagnostic group separation (Cohen' s D and area under the curve (AUC) ) of the capabilities of the methods were inspected (AD

vs . NC) . Cohen' s D is given by:

Figure imgf000035_0001
where μ and σ are mean and standard deviation. To compute the p-value for the pairwise method comparison, we carried out a two-tailed t-test for the null hypothesis of equal measures Nl-N2= 0 , where Nl and N2 were independent random measures. We computed a probability distribution for the difference between the Cohen' s D for the two measures and computed p as p (Nl > N2 ) = 1 - cdfWl _W2 (0) and p{N2 > Nl) = cdfWl _W2 (0) . The p-values for comparing the AUCs were computed using the DeLong test [14] .

As we can see in Table 1, both Cohen's D and AUC were similar in hippocampus, ventricles and medial temporal lobe measures for both the methods . However, the two-image bias correction method yielded significant improvement in group separation based on whole brain measurements .

AD NC

Mean (Std) Mean (Std) AUC Cohen' s d

WB 1.24(0.85) 0.30(0.51) 0.85(0.04) 1.34

Bias Hip -2.77(1.87) 0.84(0.04) 1.39

Correction 0.70 (0.98)

according to

the

invention

Vent 10.19(5.77) 0.81(0.05) 1.15

4.12 (4.72)

MTL -2.66(1.81) 0.86(0.04) 1.30

0.84(0.84)

WB -1.22 (0.90) 0.76(0.06) 1.07

0.37 (0.68)

Freesurfer Hip -3.37(2.01) 0.85(1.43) 0.85(0.04) 1.46 N3 Bias

Correction

Vent 10.98 (6.15) 4.39(4.84) 0.81(0.05) 1.19

MTL -2.82 (1.93) 0.85(0.04) 1.24

0.78 (1.26)

Table 1. Various statistics based on atrophy estimated using Freesufer N3 and a method according to the invention; mean and standard deviation are in % volume loss. WB : Whole

Brain, Hip: Hippocampus, Vent: Ventricles, MTL: Medial

Temporal Lobe. Bold text represents significance of two- image bias correction method over separate N3 bias

correction method.

The example of the working of the invention above uses a simultaneous registration and two-image bias correction method and demonstrates its efficiency in atrophy scoring and

diagnostic group separation capabilities. One of the key observations of the artificial example was that the

reproducibility of atrophy scores was not consistent when only the floating image was corrected for bias. This may be because of the intensity re-sampling of the floating image involved during registration. This observation implies that the choice of the image on which the bias field is modeled becomes crucial. In instances such as template-image

registration, care needs to be taken to make sure the bias is modeled on the image and the template is warped. The

artificial examples further showed that this particular choice of bias-warp is not important when the presented (bilateral) method is used. In addition, the pro- posed regularization term ensures that the bias is consistently applied to both images thereby removing any bias.

Further, single-image bias model cannot be assumed for all registration applications, for instance, cross modality registration. Whereas, due to the fact that two-image bias correction method assumes independent bias models, this method is more suited for a richer range of registration applications .

In this specification, unless expressly otherwise indicated, the word 'or' is used in the sense of an operator that returns a true value when either or both of the stated conditions is met, as opposed to the operator 'exclusive or' which reguires that only one of the conditions is met. The word 'comprising' is used in the sense of 'including' rather than in to mean 'consisting of . All prior teachings acknowledged above are hereby incorporated by reference. No acknowledgement of any prior published document herein should be taken to be an admission or representation that the teaching thereof was common general knowledge in Australia or elsewhere at the date hereof.

References

Ashburner, J., Friston, K.: Morphometry. In: Human Brain Function, second edn. Academic Press (2003)

Lewis, E.B., Fox, N.C.: Correction of differential intensity inhomogeneity in Ion- gitudinal MR images. Neurolmage 23(1) (2004) 75 - 83

Sled, J.G., Pike, G.B.: Understanding intensity non-uniformity in MRI . In: Medical Image Computing and Computer-Assisted Interventation - MICCAI'98. Volume 1496 of Lecture Notes in Computer Science. Springer Berlin

Heidelberg (1998) 614-622

Vovk, U., Pernus, F., Likar, B.: A review of methods for correction of intensity inhomogeneity in MRI . IEEE Transactions on Medical Imaging 26(3) (March 2007) 405- 421

Sled, J.G., Zijdenbos, A. P., Evans, A.C.: A nonparametric method for automatic correction of intensity nonuniformity in MRI data. IEEE Transactios on Medical Imaging 17(1) (February 1998) 87-97

Leung, K.K., Ridgway, G.R., Ourselin, S., Fox,

N.C.: Consistent multi-time-point brain atrophy estimation from the boundary shift integral. Neurolmage 59(4) (2012) 3995 - 4005

Ashburner, J., Ridgway, G.R.: Symmetric diffeomorphic modelling of longitudinal structural MRI.

Frontiers in Neuroscience 6(197) (2013)

Andersson, J., Smith, S., Jenkinson, M.: FNIRT- FMRIB's non-linear image regis- tration tool. 14th Annual Meeting of the Organization for Human Brain Mapping (2008)

Holland, D., Dale, A.M.: Nonlinear registration of longitudinal images and measureent of change in regions of interest. Medical Image Analysis 15(4) (2011) 489 - 497. Modat, M., Ridgway, G.R., Hawkes, D.J., Fox, N.C.,

Ourselin, S.: Nonrigid registration with differential bias correction using normalised mutual information. In: IEEE

International Symposium on Biomedical Imaging: From Nano to Macro. (April 2010) 356-359

. Daga, P., Modat, M. , Winston, G., White, M., Mancini,

L., McEvoy, A.W., Thornton, J., Yousry, T., Duncan, J.S.,

Ourselin, S.: Susceptibility artefact correction by

combining bO field maps and non-rigid registration using

graph cuts (2013)

. Pai, A., Darkner, S., Sorensen, L., Larsen, L.,

Mysling, P., Sporring, J., Dam, E., Nielsen, M.: Evaluation of bias in brain atrophy estimation. MICCAI 2012 Work- shop on Novel Imaging Biomarkers for Alzheimer's Disease and

Related Disorders (NIBAD'12) (2012) 198-206

. Rueckert, D., Sonoda, L.I., Hayes, C., Hill, D.L.G.,

Leach, M.O., Hawkes, D.: Non-rigid registration using free- form deformations: Application to breast MR images. IEEE

Transactions on Medical Imaging 18(8) (August 1999) 712-721

. Pai, A., Sorensen, L., Darkner, S., Mysling, P.,

Jorgensen, D., Dam, E., Lillholm, M., Oh, J., Chen, G.,

Suhy, J., Sporring, J., Nielsen, M.: Cube propagation for

focal brain atrophy estimation. In: IEEE 10th International

Symposium on Biomedical Imaging (ISBI) . (April 2013) 402-405

. Viola, P., Wells, III, W.M.: Alignment by maximization of mutual information. International Journal on Computer Vision 24(2) (September 1997) 137-154

. Mattes, D., Haynor, D., Vesselle, H., Lewellen, T.,

Eubank, W. : PET-CT im- age registration in the chest using

free-form deformations. IEEE Transactions on Medical Imaging

22(1) (January 2003) 120-128

. Liu, D.C., Nocedal, J.: On the limited memory BFGS

method for large scale optimization. Mathematical

Programming 45(3) (December 1989) 503-528

. Chong, K., Lau, W.C., Leong, J., Suhy, J., Oh, J.:

Longitudinal volumetric MRI analysis for use in alzheimer's disease multi-site clinical trials: Comparison to anal- ysis methods used in ADNI and correlation to MMSE change.

Alzheimer's & Dementia 6(4) (2010) [la] M. Miller, "Computational anatomy: shape, growth, and atrophy comparison via diffeomorphisms, " Neurolmage, vol. 23, no. 1, pp. 19-33, 2004.

[2a] M. Beg, M. Miller, A. Trouv'e, and L. Younes, "Computing large deformation metric mappings via geodesic flows of diffeomorphisms . " International Journal of Computer Vision, vol. 61, pp. 139-157, 2005.

[3a] L. Risser, F.-X. Vialard, R. Wolz, M. Murgasova, D. D. Holm, D. Rueckert, and A. D. Neuroimaging, "Simultaneous multi-scale registration using large deformation diffeomorphic metric mapping," IEEE Transactions on Medical Imaging, vol. 30, no. 10, pp. 1746- 1759, 2011.

[4a] M. Bruveris, F. Gay-Balmaz, D. D. Holm, and T. S. Ratiu, "The momentum map representation of images," Journal of Nonlinear

Science, vol. 21, no. 1, pp. 115-150, 2011.

[5a] M. De Craene, G. Piella, O. Camara, N. Duchateau, E. Silva, A. Doltra, J. D'hooge, J. Brugada, M. Sitges, and A. Frangi, "Temporal diffeomorphicfree-form deformation: Application to motion and strain estimation from 3D echocardiography, " Medical Image

Analysis, vol. 16, pp. 427-50, 2012.

[6a] A. Jain and L. Younes, "A kernel class allowing for fast computations in shape spaces induced by diffeomorphisms, " Journal of Computational and Applied Mathematics, vol. 245, pp. 162-181, 2013.

[7a] J. Ashburner, "A fast diffeomorphic image registration algorithm, "Neurolmage , vol. 38, no. 1, pp. 95-113, 2007.

[8a] M. Modat, P. Daga, M. J. Cardoso, and S. Ourselin, "Parametric nonrigid registration using a stationary velocity field, " in IEEE Workshop on Mathematical Methods in Biomedical Image Analysis (MMBIA) , 2012, pp. 145 - 150.

[9a] S. Sommer, F. Lauze, M. Nielsen, and X. Pennec, "Sparse multiscale diffeomorphic registration: the kernel bundle

framework," J. of Mathematical Imaging and Vision, vol. 46, no. 3, pp. 292-308, 2012.

[10a] A. Sotiras, C. Davatzikos, and N. Paragios, "Deformable medical image registration: a survey," IEEE Transactions on Medical Imaging, vol. 32, no. 7, pp. 1153-1190, 2013.

[11a] A. Klein, J. Andersson, B. A. Ardekani, J. Ashburner, B. B. Avants, M.-C. Chiang, G. E. Christensen, D. L. Collins, J. C. Gee, P. Hellier, J. H. Song, M. Jenkinson, C. Lepage, D. Rueckert, P. M. Thompson, T. Vercauteren, R. P. Woods, J. J. Mann, and R. V.

Parsey, "Evaluationof 14 nonlinear deformation algorithms applied to human brain mri registration," Neurolmage, pp. 786-802, 2009.

[12a] V. Arsigny, O. Commowick, X. Pennec, and N. Ayache, "A logeuclidean framework for statistics on diffeomorphisms, " in

Medical Image Computing and Computer-Assisted Intervention, vol. 4190, 2006, pp. 924-931. [13a] T. Vercauteren, X. Pennec, A. Perchant, and N. Ayache, "Non- parametric diffeomorphic image registration with the demons algorithm, " in Medical Image Computing and Computer Assisted

Intervention (MICCAI), vol. 4792, 2007, pp. 319-326.

[14a] M. Lorenzi and X. Pennec, "Geodesies, parallel transport and oneparameter subgroups for diffeomorphic image registration," International Journal of Computer Vision, vol. 105, pp. 111-127, 2013.

[15a] V. M, M. MI, Y. L, and T. A., "Statistics on diffeomorphisms via tangent space representations," Neurolmage, vol. Suppl 1, no. 23, pp. S161-9, 2004.

[16a] N. Singh, P. Fletcher, J. Preston, L. Ha, R. King, J. Marron, M. Wiener, and S. Joshi, "Multivariate statistical analysis of deformation momenta relating anatomical shape to neuropsychological measures," Medical Image Computing and Computer-Assisted

Intervention, pp. 529-537, 2010.

[17a] M. Bossa, E. Zacur, S. Olmos, and for the Alzheimer's Disease Neuroimaging Initiative., "Tensor-based morphometry with stationary velocity field diffeomorphic registration: application to adni," Neurolmage, vol. 51, no. 3, pp. 956-969, 2010.

[18a] T. Mansi, X. Pennec, M. Sermesant, H. Delingette, and N.

Ayache, "iLogDemons: A demons-based registration algorithm for tracking incompressible elastic biological tissues," Int. J.

Comput. Vis, vol. 92, no. 1, pp. 92-111, 2011.

[19a] M. Hernandez, M. N. Bossa, and S. Olmos, "Registration of anatomical images using paths of diffeomorphisms parameterized with stationary vector field flows," International Journal of Computer Vision, vol. 85, no. 3, pp. 291-306, 2009.

[20a] L. Younes, Shapes and Diffeomorphisms . Springer, 2010, vol. 171.

[21a] H. Wendland, "Piecewise polynomial, positive definite and compactly supported radial functions of minimal degree," Advances in Computational Mathematics, vol. 4, no. 1, pp. 389-396, 1995.

[22a] A. Pai, S. Sommer, L. Sorensen, S. Darkner, J. Sporring, and M. Nielsen, "Image registration using stationary velocity fields parameterized by norm-minimizing wendland kernel," in SPIE Medical Imaging, 2015, (to appear in print) .

[23a] M. Bossa, E. Zacur, and S. Olmos, "Algorithms for computing the group exponential of diffeomorphisms : Performance evaluation," Computer Vision and Pattern Recognition Workshops, pp. 1 - 8, 2008.

[24a] G. E. Christensen and H. J. Johnson, "Consistent image registration, "IEEE Transactions on Medical Imaging, vol. 20, pp. 568-582, 2001.

[25a] A. Leow, S. Huang, A. Geng, J. Becker, S. Davis, A. Toga, and P. Thompson, "Inverse consistent mapping in 3d deformable image registration: its construction and statistical properties," in Information Processing in Medical Imaging, vol. 19, 2005, pp. 493- 503.

[26a] J. Modersitzki , Numerical methods for image registration. Oxford University Press, 2004.

[27a] Q. Ye, "Reproducing kernels of generalized sobolev spaces via a green function approach with differential operators," Illinois Institute of Technology, Tech. Rep., 2010.

[28a] N. Aronszajn, "Theory of reproducing kernels," Transactions of the American Mathematical Society, vol. 68, no. 3, pp. 337-404, 1950.

[29a] M. Fornefett, K. Rohr, and H. Stiehl, "Radial basis functions with compact support for elastic registration of medical images," Image and Vision Computing, vol. 19, no. 1, pp. 87-96, 2001.

[30a] see [17]

[31a] A. Pai, L. Sorensen, S. Darkner, P. Mysling, D. Jorgensen, E. Dam, M. Lillholm, J. Oh, G. Chen, J. Suhy, J. Sporring, and M.

Nielsen, "Cube propagation for focal brain atrophy estimation," in IEEE symposium on biomedical imaging, 2013.

[32a] B. Wyman, D. Harvey, K. Crawford, M. Bernstein, O.

Carmichael, P. Cole, P. Crane, C. Decarli, N. Fox, J. Gunter, D.

Hill, R. Killiany, C. Pachai, A. Schwarz, N. Schuff, M. Senjem, J. Suhy, P. Thompson, M. Weiner, C. J. Jack, and A. D. N. Initiative, "Standardization of analysis sets for reporting results from ADNI MRI data," Alzheimer's and Dementia, October 2012.

[33a] A. C. Evans, D. L. Collins, and B. Milner, "An MRI-based stereotactic brain atlas from 300 young normal subjects,"

Proceedings of the 22nd Symposium of the Society for Neuroscience , p. 408, 1992.

[34a] K. J. Friston, J. Ashburner, C. D. Frith, J.-B. Poline, J. D. Heather, and R. S. J. Frackowiak, "Spatial registration and normalization of images, "Human Brain Mapping, vol. 3, no. 3, pp. 165-189, 1995.

[35a] A. Nieto-Castanon, S. S. Ghosh, J. A. Tourville, and F. H. Guenther, "Region of interest based analysis of functional imaging data," Neurolmage, vol. 19, no. 4, pp. 1303 - 1316, 2003.

[36a] J. Tourville and F. Guenther, "A cortical and cerebellar parcellation system for speech studies," Boston University

Technical Reports, Tech. Rep. CAS/CNS -03-022 , 2003.

[37a] A. Dale, B. Fischl, and M. Sereno, "Cortical surface -based analysis. I . Segmentation and surface reconstruction," Neuroimage, vol. 9, pp. 179-194, 1999.

[38a] B. Zhou, A. Pai, and M. Nielsen, "Simultaneous registration and bilateral differential bias correction in brain mri," IntellMR, MICCAI workshop, 2014. [39a] X. Hua, D. P. Hibar, C. R. Ching, C. P. Boyle, P.

Rajagopalan, B. A. Gutman, A. D. Leow, A. W. Toga, C. R. J. Jr., D.

Harvey, M. W. Weiner, and P. M. Thompson, "Unbiased tensor-based morphometry: Improved robustness and sample size estimates for alzheimer's disease clinical trials," Neuroimage, vol. 66, pp. 648- 661, 2013.

[40a] B. Efron and R. J. Tibshirani, An introduction to the bootstrap. Chapman and Hall, New York, 1993.

[41a] D. Holland, L. McEvoy, A. Dale, and A. D. N. Initiative., "Unbiased comparison of sample size estimates from longitudinal structural measures in ADNI," Human Brain Mapping, vol. 33, no. 11, pp. 2586-2602, 2012.

[42a] E. R. DeLong, D. M. DeLong, and D. L. Clarke -Pearson ,

"Comparing the Areas under Two or More Correlated Receiver

Operating Characteristic Curves: A Nonparametric Approach,"

Biometrics, vol. 44, no. 3, pp. 837-845, Sep. 1988.

[43a] C. Jack, R. Petersen, Y. Xu, P. O'Brien, G. Smith, I. RJ, T. EG, and K. E., "Rate of medial temporal lobe atrophy in typical aging and alzheimer's disease," Neurology, vol. 51, no. 4, pp. 993- 999, 1998.

[44a] M. Hernandez, M. N. Bossa, and S. Olmos, "Registration of anatomical images using paths of diffeomorphisms parameterized with stationary vector field flows," International Journal of Computer Vision, vol. 85, pp. 291-306, 2009.

[45a] C. Chefd'hotel, G. Hermosillo, and O. Faugeras, "Flows of diffeomorphisms for multimodal image registration," in IEEE

International Symposium on Biomedical Imaging, 2002, pp. 753-756.

[46a] M. Bro-Nielsen and C. Gramkow, "Fast fluid registration of medical images," International Conference on Visualization in Biomedical Computing, pp. 267-276, 1996.

[47a] N. Navab, A. Kamen, and D. Zikic, "Unifying characterization of deformable registration methods based on the inherent

parametrization : An attempt at an alternative analysis approach," 4th International Workshop on Biomedical Image Registration, WBIR 2010, vol. 6204, pp. 161-172, 2010.

[48a] X. Pennec, P. Cachier, and N. Ayache, "Understanding the demon's algorithm: 3d non-rigid registration by gradient descent," in Medical Image Computing and Computer-Assisted Intervention, vol. 1679, 1999, pp. 597-605.

[49a] M. Nielsen, L. Florack, and R. Deriche, "Regularization, scale-space , and edge detection filters," ECCV, vol. 1065, pp. 70- 81, 1996.

[50a] W. Shi, M. Jantsch, P. Aljabar, L. Pizarro, W. Bai, H. Wang, D. 0'Regan,X. Zhuang, and D. Rueckert, "Temporal sparse free-form deformations, "Medical Image Analysis, vol. 17, no. 7, pp. 779-789, 2013. [51a] E. Haber and J. Modersitzki , "A multilevel method for image registration, "SIAM Journal on scientific computing, vol. 27, no. 5, pp. 1594-1607, 2006.

[52a] W. Cai and J. Wang, "Adaptive multiresolution collocation methods forinitial boundary value problems of nonlinear pdes," SIAM J. Numer. Anal., vol. 33, no. 3, pp. 937-970, 1996.

Claims

A method of computer analysis of data sets representing images to achieve bias correction and image registration, each image including a bias in intensity within the image of unknown magnitude, the method comprising:
a) inputting a digital data set of a first image and a digital data set of a second image into a computer;
b) in said computer calculating a deformation of said first
image that transforms said first image into a transformed image that is an optimised approximation of said second image and
c) simultaneously calculating and applying a bias correction which is applied to said first image and a bias correction which is applied to said transformed image such that each of the first image and the transformed image is individually corrected for bias therein.
A method as claimed in claim 1, wherein an average of the bias correction over the first image is equal and opposite to an average of the bias correction over said transformed image.
A method as claimed in claim lor claim 2, wherein each said data set represents an MRI image.
A method as claimed in claim 3, wherein said MRI image is of a brain or a part thereof.
A method as claimed in any preceding claim, wherein each said data set represents an image of the same object from the same viewpoint obtained at spaced time points .
A method as claimed in any one of claims 1 to 4 , wherein each said data set represents an image of a different instance of the same type of object from the same viewpoint.
A method as claimed in claim 6, wherein one said image is adopted as fiducial and each of a pair of other said images is subjected to bias correction and registration with said fiducial image, whereby to better define the bias correction for said fiducial image.
A method as claimed in claim 6, wherein multiple images are bias corrected and registered to produce a new image which is ί bias corrected centroid of the multiple images.
A method as claimed in any preceding claim, wherein
simultaneous bias correction and said deformation calculation is conducted by optimising an objective function which
comprises a term which is minimised when the images are optimally registered.
A method as claimed in claim 9, wherein said objective function further comprises a regularisation term which
penalises deviation from linearity of calculated parameterised bias and deformation fields.
A method as claimed in claim 9 or claim 10, wherein said objective function further comprises a term constraining the average bias applied to each image to be essentially equal and opposite .
A method as claimed in claim 5, wherein from the bias corrected and registered images an estimate of change of size of said object between said time points is obtained.
A computer programmed to accept as input data sets
representing images and to process said data sets to achieve bias correction and image registration, each image including a bias in intensity within the image of unknown magnitude, said processing in said computer comprising
calculating a deformation of said first image that transforms said first image into a transformed image that is an optimised approximation of said second image and
simultaneously calculating and applying a bias correction whicl is applied to said first image and a bias correction which is applied to said transformed image such that each of the first image and the transformed image is individually corrected for bias therein.
PCT/EP2015/071248 2014-09-17 2015-09-16 Bias correction in images WO2016042037A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
GB1416416.4 2014-09-17
GBGB1416416.4A GB201416416D0 (en) 2014-09-17 2014-09-17 Bias correction in images

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US15/511,770 US10332241B2 (en) 2014-09-17 2015-09-16 Bias correction in images

Publications (1)

Publication Number Publication Date
WO2016042037A1 true WO2016042037A1 (en) 2016-03-24

Family

ID=51869739

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2015/071248 WO2016042037A1 (en) 2014-09-17 2015-09-16 Bias correction in images

Country Status (3)

Country Link
US (1) US10332241B2 (en)
GB (1) GB201416416D0 (en)
WO (1) WO2016042037A1 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018091360A1 (en) 2016-11-17 2018-05-24 Koninklijke Philips N.V. Intensity corrected magnetic resonance images
WO2018098213A1 (en) * 2016-11-23 2018-05-31 Wake Forest University Health Sciences Medical image analysis using mechanical deformation information

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2009058915A1 (en) * 2007-10-29 2009-05-07 The Trustees Of The University Of Pennsylvania Computer assisted diagnosis (cad) of cancer using multi-functional, multi-modal in-vivo magnetic resonance spectroscopy (mrs) and imaging (mri)
US20130315448A1 (en) * 2012-03-28 2013-11-28 Evan Fletcher Systems and methods for measuring longitudinal brain change incorporating boundary-based analysis with tensor-based morphometry

Family Cites Families (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US100190A (en) * 1870-02-22 Improvement in wagon-seats
FR2763721B1 (en) * 1997-05-21 1999-08-06 Inst Nat Rech Inf Automat Electronic device for processing images for the detection of dimensional variations
WO2006114003A1 (en) * 2005-04-27 2006-11-02 The Governors Of The University Of Alberta A method and system for automatic detection and segmentation of tumors and associated edema (swelling) in magnetic resonance (mri) images
US20070237372A1 (en) * 2005-12-29 2007-10-11 Shoupu Chen Cross-time and cross-modality inspection for medical image diagnosis
US20070249928A1 (en) * 2006-04-19 2007-10-25 General Electric Company Method and system for precise repositioning of regions of interest in longitudinal magnetic resonance imaging and spectroscopy exams
US7813592B2 (en) * 2006-08-09 2010-10-12 Siemens Medical Solutions Usa, Inc. System and method for non-rigid multi-modal registration on the GPU
US7961925B2 (en) * 2006-11-14 2011-06-14 Siemens Aktiengesellschaft Method and system for dual energy image registration
US7778488B2 (en) * 2007-03-23 2010-08-17 Varian Medical Systems International Ag Image deformation using multiple image regions
WO2009038822A2 (en) * 2007-05-25 2009-03-26 The Research Foundation Of State University Of New York Spectral clustering for multi-type relational data
DE102007033897B4 (en) * 2007-07-20 2010-02-11 Siemens Ag Method for correcting distortions in image data records recorded by means of a magnetic resonance apparatus and computer program for carrying out this method
US7535227B1 (en) * 2007-10-26 2009-05-19 General Electric Company Method and apparatus for correcting distortion in MR images caused by metallic implants
GB0913930D0 (en) * 2009-08-07 2009-09-16 Ucl Business Plc Apparatus and method for registering two medical images
JP5612371B2 (en) * 2010-06-11 2014-10-22 富士フイルム株式会社 Image alignment apparatus and method, and program
US8942512B2 (en) * 2011-12-24 2015-01-27 Ecole De Technologie Superieure Methods and systems for processing a first image with reference to a second image
WO2013136278A1 (en) * 2012-03-15 2013-09-19 Koninklijke Philips N.V. Multi-modality deformable registration
US8923652B2 (en) * 2012-10-25 2014-12-30 Nvidia Corporation Methods and apparatus for registering and warping image stacks
GB201301795D0 (en) * 2013-02-01 2013-03-20 Ucl Business Plc Apparatus and method for correcting susceptibility artefacts in a magnetic resonance image
CN104036452B (en) * 2013-03-06 2017-12-05 东芝医疗系统株式会社 Image processing apparatus and method and medical image equipment
JP6505078B2 (en) * 2013-03-29 2019-04-24 コーニンクレッカ フィリップス エヌ ヴェKoninklijke Philips N.V. Image registration
US20160143576A1 (en) * 2013-07-15 2016-05-26 Tel Hashomer Medical Research Infrastructure And Services Ltd. Mri image fusion methods and uses thereof
CA2932259A1 (en) * 2013-12-03 2015-06-11 Viewray Technologies, Inc. Single- and multi-modality alignment of medical images in the presence of non-rigid deformations using phase correlation
CN106030653B (en) * 2014-02-24 2019-04-05 华为技术有限公司 For generating the image processing system and image processing method of high dynamic range images
KR20150118484A (en) * 2014-04-14 2015-10-22 삼성전자주식회사 Method and Apparatus for medical image registration
US10062167B2 (en) * 2014-08-15 2018-08-28 Toshiba Medical Systems Corporation Estimated local rigid regions from dense deformation in subtraction
US10083506B2 (en) * 2014-11-07 2018-09-25 Antaros Medical Ab Whole body image registration method and method for analyzing images thereof
US9760983B2 (en) * 2015-10-19 2017-09-12 Shanghai United Imaging Healthcare Co., Ltd. System and method for image registration in medical imaging system

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2009058915A1 (en) * 2007-10-29 2009-05-07 The Trustees Of The University Of Pennsylvania Computer assisted diagnosis (cad) of cancer using multi-functional, multi-modal in-vivo magnetic resonance spectroscopy (mrs) and imaging (mri)
US20130315448A1 (en) * 2012-03-28 2013-11-28 Evan Fletcher Systems and methods for measuring longitudinal brain change incorporating boundary-based analysis with tensor-based morphometry

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
JABER JUNTU ET AL: "Bias Field Correction for MRI Images", COMPUTER RECOGNITION SYSTEMS - ADVANCES IN SOFT COMPUTING, vol. 30, 22 May 2005 (2005-05-22), pages 543, XP055023554, DOI: 10.1007/3-540-32390-2_64 *
MAES F ET AL: "MEDICAL IMAGE REGISTRATION USING MUTUAL INFORMATION", PROCEEDINGS OF THE IEEE, IEEE. NEW YORK, US, vol. 91, no. 10, October 2003 (2003-10-01), pages 1699 - 1722, XP008056001, ISSN: 0018-9219, DOI: 10.1109/JPROC.2003.817864 *
MAINTZ J B A ET AL: "Mutual-information-based registration of medical images: a survey", IEEE TRANSACTIONS ON MEDICAL IMAGING, IEEE SERVICE CENTER, PISCATAWAY, NJ, US, vol. 22, no. 8, August 2003 (2003-08-01), pages 986 - 1004, XP011099100, ISSN: 0278-0062, DOI: 10.1109/TMI.2003.815867 *
MARC MODAT ET AL: "Nonrigid registration with differential bias correction using normalised mutual information", BIOMEDICAL IMAGING: FROM NANO TO MACRO, 2010 IEEE INTERNATIONAL SYMPOSIUM ON, IEEE, PISCATAWAY, NJ, USA, 14 April 2010 (2010-04-14), pages 356 - 359, XP031693609, ISBN: 978-1-4244-4125-9 *
URO VOVK ET AL: "A Review of Methods for Correction of Intensity Inhomogeneity in MRI", IEEE TRANSACTIONS ON MEDICAL IMAGING, IEEE SERVICE CENTER, PISCATAWAY, NJ, US, vol. 26, no. 3, March 2007 (2007-03-01), pages 405 - 421, XP011171979, ISSN: 0278-0062, DOI: 10.1109/TMI.2006.891486 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018091360A1 (en) 2016-11-17 2018-05-24 Koninklijke Philips N.V. Intensity corrected magnetic resonance images
WO2018098213A1 (en) * 2016-11-23 2018-05-31 Wake Forest University Health Sciences Medical image analysis using mechanical deformation information

Also Published As

Publication number Publication date
GB201416416D0 (en) 2014-10-29
US20170243336A1 (en) 2017-08-24
US10332241B2 (en) 2019-06-25

Similar Documents

Publication Publication Date Title
Kybic et al. Fast parametric elastic image registration
Mattes et al. PET-CT image registration in the chest using free-form deformations
Chung et al. Deformation-based surface morphometry applied to gray matter deformation
Despotović et al. MRI segmentation of the human brain: challenges, methods, and applications
Avants et al. A reproducible evaluation of ANTs similarity metric performance in brain image registration
Myronenko et al. Intensity-based image registration by minimizing residual complexity
Smistad et al. Medical image segmentation on GPUs–A comprehensive review
Ardekani et al. Quantitative comparison of algorithms for inter-subject registration of 3D volumetric brain MRI scans
Van Ginneken et al. Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database
Periaswamy et al. Medical image registration with partial data
Rohlfing et al. Volume-preserving nonrigid registration of MR breast images using free-form deformation with an incompressibility constraint
Yang et al. Quicksilver: Fast predictive image registration–a deep learning approach
US6985612B2 (en) Computer system and a method for segmentation of a digital image
Caldairou et al. A non-local fuzzy segmentation method: application to brain MRI
Penney et al. Registration-based interpolation
Heinrich et al. MIND: Modality independent neighbourhood descriptor for multi-modal deformable registration
US8682054B2 (en) Method and system for propagation of myocardial infarction from delayed enhanced cardiac imaging to cine magnetic resonance imaging using hybrid image registration
US7502499B2 (en) System and method for filtering noise from a medical image
US6813373B1 (en) Image segmentation of embedded shapes using constrained morphing
Liu et al. Deformable registration of cortical structures via hybrid volumetric and surface warping
Rueckert et al. Nonrigid registration of medical images: Theory, methods, and applications [applications corner]
Trinh et al. Novel example-based method for super-resolution and denoising of medical images
Wang et al. Automatic segmentation of neonatal images using convex optimization and coupled level sets
Periaswamy et al. Elastic registration in the presence of intensity variations
Hernandez et al. Registration of anatomical images using paths of diffeomorphisms parameterized with stationary vector field flows

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15766136

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15511770

Country of ref document: US

NENP Non-entry into the national phase in:

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 04.07.2017)

122 Ep: pct application non-entry in european phase

Ref document number: 15766136

Country of ref document: EP

Kind code of ref document: A1