Geometry of Diffeomorphism Groups and Shape Matching
Imperial College London Department of Mathematics
Geometry of Diffeomorphism Groups and Shape Matching
Martins Bruveris
May 31, 2012
Supervised by Prof. Darryl D. Holm
Submitted in part fulfilment of the requirements for the degree of Doctor of Philosophy in Mathematics of Imperial College London and the Diploma of Imperial College London
Declaration
I herewith certify that all material in this dissertation which is not my own work has been duly acknowledged. Selected results from this dissertation have been disseminated in scientific publications as detailed in Section 1.4.
Martins Bruveris
3
Abstract
The large deformation matching (LDM) framework is a method for registra- tion of images and other data structures, used in computational anatomy. We show how to reformulate the large deformation matching framework for registration in a geometric way. The general framework also allows to gen- eralize the large deformation matching framework to include multiple scales by using the iterated semidirect product of groups. An important ingredient in the LDM framework is the choice of a suitable Riemannian metric on the space of diffeomorphisms. Since the space in question is infinite-dimensional, not every choice of the metric is suitable. In particular the geodesic distance, which is defined as the infimum over the length of all paths connecting two points, may vanish. For the family s of Sobolev-type H -metrics on the diffeomorphism groups of R and S1 we establish that the geodesic distance vanishes for metrics of order 0 s 1 . ≤ ≤ 2 The geodesic distance also vanishes for the L2-metric on the Virasoro-Bott group, which is a central extension of the diffeomorphism group of the circle. Vanishing of geodesic distance implies that the length-functional, which assigns to each curve in the manifold its length, has no global minima, when restricted to paths with fixed endpoints. We show that for the L2-metric on the diffeomorphism group of R and the Virasoro-Bott group doesn’t have any local minima either. The large deformation matching framework is not the only approach to the registration and shape comparison. For curves and surfaces it is possible to define a Riemannian metric directly on the space of curves or surfaces and use geodesics with respect to this metric to measure differences in shape. We use the family of Sobolev-type metrics on surfaces from [7]. We show how to discretize the geodesic equations and solve the boundary value problem via a shooting method on the initial velocity. The discrete equations are implemented via the finite element method.
5
To my family.
7
Acknowledgments
Over the last three years I had the fortune to meet many people, to whom I owe my gratitude and without whom this work would not have been possible. First I want to thank my adviser Darryl Holm, who gave me the opportunity to study in London and guided me on my first steps in the world of research. I also want to thank Peter Michor for his valuable advice, both technical and personal and for always making me feel welcome in Vienna. It is due to my colleagues in London, David Ellis, Laurent Risser, Sehun Chun, Fran¸cois- Xavier Vialard, David Meier, Chris Cantwell and Christopher Burnett, that I always felt like part of a team and never had to look far for someone to drink coffee with. It was a pleasure to collaborate with my friends in Vienna, Martin Bauer and Philipp Harms. My thanks also goes to Colin Cotter, Fran¸coisGay-Balmaz and Tudor Ratiu for interesting discussions and valuable advice. And finally I want to thank all the members of the ShapeFRG meetings, who made working in this field like being part of a large family.
9
Contents
Abstract 5
Acknowledgments 9
1 Introduction 13 1.1 Diffeomorphism Group and Applications ...... 13 1.2 Content of this Work ...... 15 1.3 Contributions of this Work ...... 17 1.4 Publications ...... 18
2 Multiscale Registration 19 2.1 Geometry of Registration ...... 19 2.1.1 Motivation ...... 19 2.1.2 Abstract Framework ...... 24 2.2 Registration using Diff(Ω) ...... 31 2.2.1 The Setting ...... 31 2.2.2 Matching Problems ...... 34 2.2.3 Landmark Matching ...... 36 2.2.4 Image Matching ...... 37 2.2.5 Vector Field Matching ...... 39 2.3 Multiscale Registration ...... 41 2.3.1 Semidirect Products ...... 42 2.3.2 Semidirect Products of Diffeomorphism Groups . . . . 45 2.3.3 Sums of Kernels ...... 49 2.3.4 The Order Reversed ...... 53 2.3.5 A Continuum of Scales ...... 56 2.3.6 Restriction to a Finite Number of Scales ...... 60
3 Geodesic Distance on Diffeomorphism and Related Groups 63 3.1 Overview of the Results ...... 63
11 3.2 Mathematical Background ...... 65 3.2.1 Diffeomorphism Groups ...... 65 3.2.2 Sobolev Spaces on Manifolds ...... 66
3.2.3 Sobolev Metrics on Diffc(M)...... 68 3.2.4 Virasoro-Bott Group ...... 69 3.3 Diffeomorphism Groups ...... 70 3.4 Virasoro-Bott Group ...... 82 3.5 Local Minima of the Length Functional ...... 89 3.6 Outlook ...... 100
4 Surface Matching 103 4.1 Background ...... 103 4.2 First Order Sobolev-type Metric ...... 104 4.3 Variation of the Metric ...... 105 4.4 The Helmholtz Operator and Duality ...... 110 4.5 Discretization ...... 113 4.5.1 The Geodesic Equation ...... 113 4.5.2 Computing the Gradient ...... 116 4.6 Numerical Experiments ...... 120 4.7 Outlook ...... 124
Bibliography 127
12 1 Introduction
1.1 Diffeomorphism Group and Applications
The diffeomorphism group plays a central role in the field of computational anatomy. For the first time its use in biology was noted by D’Arcy Thomp- son in his book “On Growth and Form” [79]. Following the paradigm of pattern theory, introduced by Grenander [31, 33], we assume that the anatomical variety of an object of interest across the population can be explained by choosing a template, upon which a set of deformations acts, thus generating the entire population, potentially up to small-scale noise. In this setting the collection of anatomical objects forms a homogeneous space under the group of deformations [32]. The objects, which are stud- ied in computational anatomy are those that can be obtained by medical imaging procedures. They include volumetric gray-level images from MRI and CT scans, vector and tensor fields from diffusion tensor MRI, surfaces representing the outline of organs, curves in space representing white mat- ter fiber tracts and manually or automatically assigned feature points. The group of deformations is usually the diffeomorphism group of the ambient three-dimensional space. The aim of the pattern theory approach is to encode the differences be- tween two objects in the deformation that matches one object to the other. This allows one to ignore all the complexity inherent in anatomical struc- tures and only concentrate on the differences between them. Therefore, when given a template object I V and a target object I V from 0 ∈ targ ∈ the population, denoted by V and assumed to be a vector space, as a first step, it is necessary to determine the deformation g G from the group G ∈ of all deformations, mapping I0 to g.I0 = Itarg. This can be done using the large deformation diffeomorphic metric matching (LDM) framework of [40, 10, 61, 91, 93]. The objective of LDM is not just to determine a deformation g G such 1 ∈
13 that the group action g I of g G on the template I V approximates 1 0 1 ∈ 0 ∈ the target I V to within a certain tolerance. Rather, the objective of targ ∈ LDM is to find the optimal path gt G continuously parametrized by time ∈ t R that smoothly deforms I0 through It = gtI0 to g1I0. The optimal path ∈ gt G is defined as the path that costs the least in time-integrated kinetic ∈ energy for a given tolerance. Hence, the deformable template method may be formulated as an optimization problem based on a trade-off between the following two properties: (1) the tolerance for inexact matching between the final deformed template g1I0 and the target template Itarg; and (2) the cost of time-integrated kinetic energy of the rate of deformation along the path gt. The former is defined by assigning a norm : V R to measure k · k → the mismatch gtI I between the two images. The latter is obtained k 0 − targk by choosing a Riemannian metric : TG R that defines the kinetic | · | → energy on the tangent space TG of the group G. In applications of LDM to the analysis of features in bio-medical images, the optimal path gt is naturally chosen from among the diffeomorphic trans- formations G = Diff(Ω) of an open, bounded domain Ω. The domain Ω will be taken to be the ambient space in which the anatomy is located. It can be shown that [64, 89] the optimal path of deformation satisfies an evolution equation, which is the geodesic equation for the metric, which was chosen to measure the kinetic energy. Therefore the whole path is encoded in its initial value at time t = 0. These geodesic equations on the diffeomorphism group naturally lead us the field of hydrodynamics [38]. The importance of the diffeomorphism group was also realized in a differ- ent context. In the seminal work of Arnold [3] it was realized that Euler’s equations, which govern the motion of incompressible fluids, can be inter- preted as geodesic equations on the group of volume-preserving diffeomor- phisms with respect to the right-invariant L2 metric. This interpretation was used by Ebin and Marsden [23] to rewrite Euler’s equation, a system of PDEs, as a second order ODE on a suitable Hilbert space and thus to prove the well-posedness of Euler’s equations in three dimensions. Inspired by Arnold, it was shown that other PDEs of hydrodynamic type can be interpreted as geodesic equations on diffeomorphism or related groups with respect to certain Riemannian metrics: examples include Burgers’ equation for the diffeomorphism group with respect to the L2-metric, the Camassa-Holm equation with respect to the H1-metric [14, 19, 28] and the
14 KdV equation, which is a geodesic equation for the Virasoro-Bott group with respect to the L2-metric [65, 73]. Closely associated are also the Hunter-Saxton and the modified Constantin-Lax-Majda equations, which arise, when the diffeomorphism group is endowed with the homogeneous H˙ 1 or H˙ 1/2-metrics [43, 88]. The interpretation of a PDE as the geodesic equation on an infinite di- mensional manifold opened up a variety of geometrical questions, which may be asked about the manifold. What is the curvature of this mani- fold? Do geodesics have conjugate points? Do there exist totally geodesic submanifolds? What are the properties of the metric induced by geodesic distance? Some of these questions have been addressed in the past. See e.g. the papers [63] for properties of conjugate points, [57, 56] for properties of geodesic distance and [42, 7, 54] for properties of the curvature tensor. The problem of matching objects and comparing shape has a history that goes beyond LDM or the field of computational anatomy. Beginning with [41], shape is understood as the properties of objects up to a group action. For example [41] studied the manifold of triangles modulo transla- tions, rotations and scalings and endowed it with a Riemannian structure. Another widely studied shape space is the space of closed unparametrized curves. Various Riemannian metrics have been proposed on this space [57, 58, 90, 74, 78, 21] and used for shape comparison. More recently atten- tion has turned to the space of surfaces embedded in R3, both parametrized [45, 49] and unparametrized [87, 48, 7]. We will concentrate in particular on the space of parametrized surfaces and the family of metrics for these, which where proposed in [7]. In contrast to the LDM framework, where a group of deformations is assumed to act on the surface and thereby induces a Riemannian metric, in this setting the metric is defined directly on the tangent space of the manifold of surfaces and takes into account the local geometry of the surface. The theoretical properties of these metrics were studied in depth in [7], but robust numerical methods for matching have yet to be developed.
1.2 Content of this Work
This work consists of three parts. In the first part we study the LDM frame- work, which is used for matching images [10], vector fields [16], landmarks
15 [40] and other data structures. We show how to formulate it abstractly in terms of a Lie group of deformations acting on a vector space of anatomical objects and derive the matching equations for them. All the above men- tioned examples can be seen as special cases of this general framework. More importantly this generalization allows us to consider other groups than the diffeomorphism group for matching. We use the freedom of changing the group of deformation to replace the diffeomorphism group by a semidirect product of diffeomorphism groups to account for multiple length-scales in the images. Other approaches to incor- porate multiscale aspects in LDM were [67, 68, 76]. We show the equivalence of all three methods and in doing so give a geometric interpretation to the sum-of-kernels strategy presented in [68]. All three approaches can also be generalized from a finite, discrete set of scales to the case of a continuum of scales. The second part deals with properties of the geodesic distance on the dif- feomorphism and Virasoro-Bott group. The geodesic distance between two points is defined as the infimum of the path-length over all paths connecting the two points. In finite dimensions, because of the local invertibility of the exponential map, this distance is always positive and the topology of the resulting metric space is the same as the manifold topology. However, in in- finite dimensions, this does not always hold: the induced geodesic distance for weak Riemannian metrics on infinite dimensional manifolds may van- ish. This surprising fact was first noticed for the L2-metric on shape space Imm(S1, R2)/ Diff(S1) in [57, 3.10]. Here Imm(S1, R2)/ Diff(S1) denotes the orbifold of all immersions Imm(S1, R2) of S1 into R2 modulo reparametriza- tions. In [56] it was shown that this result holds for the general shape space Imm(M,N)/ Diff(M) for any compact manifold M and Riemannian mani- fold N, and also for the right invariant L2-metric (or equivalently Sobolev- type metric of order zero) on each full diffeomorphism group with compact support Diffc(N). In particular, since Burgers’ equation is related to the 2 1 geodesic equation of the right invariant L -metric on Diffc(R ), it implies that solutions of Burgers’ equation are critical points of the length func- tional, but they are not length-minimizing. If the metric is too weak and the geodesic distance on the diffeomorphism group vanishes then the matching problem using this metric will be ill- posed, i.e. the minimum won’t be attained. Conversely, it was shown in
16 [56] that for Sobolev-type metrics on the diffeomorphism group of order one or higher the induced geodesic distance is positive. This naturally leads to the question whether one can determine the Sobolev order where this change of behavior occurs. We give a complete answer to this question in the case of M = S1 and a partial answer for M = R. Furthermore we show that the geodesic distance vanishes on the Virasoro-Bott group with respect to the L2-metric and that the path-length functional does not have any local minima. In the third part we explore alternative approaches to LDM for matching surfaces. The family of Sobolev-type metrics on surfaces, which was studied in [7], is a natural starting point to develop numerical methods for the matching problem of parametrized surfaces. In this work we concentrate on the Sobolev metric of first order and discretize the geodesic and matching equations using finite elements.
1.3 Contributions of this Work
The LDM matching framework is formulated in an abstract setting • thus exposing its geometric structure and the role played by the mo- mentum map. We show, that the standard examples of image, vector field and landmark matching are special cases of this abstract setting.
We show the equivalence between matching images with a semidirect • product of diffeomorphism groups and matching with a single group, but with a different metric. By letting each group in the semidirect product represent one scale, we obtain a method for multiscale match- ing.
This equivalence is then generalized from a finite number of discrete • scales to the case of a continuum of scales and show how to extract the information corresponding to each scale and how matching with a finite number of scales can be seen as a special case of the continuum of scales.
We show that the geodesic distance vanishes on the group Diffc(R) • of compactly supported diffeomorphisms of R, when equipped with a right-invariant Sobolev metric of order 0 s < 1 . ≤ 2
17 We shown that the geodesic distance also vanishes for s = 1 on the • 2 1 1 1 group Diff(S ) and that it does not vanish for s > 2 on either Diff(S ) or Diffc(R).
Furthermore it is shown that the distance vanishes on the Virasoro- • Bott group, when equipped with the right-invariant L2-distance.
In addition to knowing that the geodesic distance vanishes, we also • show that the length functional has not even local minima for either Diff(R) or the Virasoro-Bott group with the L2-metric.
A discretization for the H1-metric on the space of parametrized sur- • faces is proposed and it is shown how to solve the initial and boundary value problem for geodesics. This discretization is implemented and tested on a set of synthetic examples.
1.4 Publications
Several results from this thesis have been disseminated in scientific publica- tions, details of which are given below. The abstract formulation of the LDM framework and first results on the semidirect product (Chapter 2) were developed in collaboration with Dar- ryl D. Holm, Fran¸coisGay-Balmaz and Tudor Ratiu and published in [12]. This work was continued with Fran¸cois-Xavier Vialard and Laurent Risser to include a description of a continuum of scales, provide a variational in- terpretation of the mixture of kernels and was published in [13]. The vanishing of the geodesic distance for the L2-metric on the Virasoro- Bott group was joint work with Martin Bauer, Philipp Harms and Peter Michor and was published in [8]. The results on the non-existence of local minima was a continuation of this work and was published in [11]. The results on the Sobolev-type metrics on the diffeomorphism groups were also obtained together with Martin Bauer, Philipp Harms and Peter Michor and were published in [6]. They constitute Chapter 3. The discretization of the geodesic equation for surfaces and the match- ing problem was developed together with Martin Bauer and presented at the MICCAI 2011 workshop Mathematical Foundations of Computational Anatomy [5].
18 2 Multiscale Registration
2.1 Geometry of Registration
2.1.1 Motivation
The optimal solution to a non-rigid template matching problem is defined as the shortest, or least expensive, path of continuous deformations of one geometric object (template) into another one (target). The goal is to the find the path of deformations that is the shortest, or costs the least, for a given tolerance in matching the target. The approach focuses its attention on the properties of the action of a Lie group G of transformations on the set of deformable templates. The attribution of a cost to this process is based on metrics defined on the tangent space TG of the group G, following the principles in [31].
Formulation of LDM
In the Large Deformation Diffeomorphism Metric Matching (LDM) frame- work the template matching procedure for image registration is formulated as follows. Suppose an image, say a medical image, is acquired using MRI, CT, or some other imaging technique. To begin, consider the case that the information in an image can be represented as a function I :Ω R, where d → Ω R is the domain of the image. The set of images V = (Ω) is the ⊆ F vector space of smooth functions. One usually deals with planar (d = 2) or volumetric (d = 3) images. Consider the comparison of two images, the template image I0 and the target image I1. The goal is to find a transfor- mation ϕ :Ω Ω, such that the transformed image I ϕ−1 matches the → 0 ◦ target image I1 up to prescribed accuracy, as measured by, say, the squared L2-norm of their difference,
−1 2 E (I ,I ) = I ϕ I 2 . 2 0 1 k 0 ◦ − 1kL
19 For this purpose, one introduces a time-indexed deformation process, that starts at time t = 0 with the template (denoted I0), and reaches the target
I1 at time t = 1. At a given time t during this process, the current object It is assumed to be the image of the template, I0, obtained through a sequence of deformations. We also want the time-indexed transformation to be regular. To ensure its regularity, we require the transformation to be generated as the flow of a smooth, time dependent vector field u : [0, 1] Ω Ω, i.e. ϕ = ϕ with × → 1
∂tϕt = ut ϕt, ϕ (x) = x . (2.1) ◦ 0
We measure the regularity of ut via a kinetic-energy like term
Z 1 2 Ekin(ut) = ut H dt , 0 | | where ut H is a norm on the space of vector fields on Ω defined in terms of | | a positive self-adjoint differential operator L by
2 ut = u, Lu 2 . (2.2) | |H h iL
A possible choice for L is the Helmholtz operator Lu = u α2∆u. We − denote by the space of vector fields, for which this norm is finite. H Following [10] we can cast the problem of registering I0 to I1 as a varia- tional problem. Namely, we seek to minimize the cost
Z 1 2 1 −1 2 E(ut) = ut H dt + 2 I0 ϕ1 I1 L2 (2.3) 0 | | 2σ k ◦ − k over all time-dependent vector fields ut. The transformation ϕ1 is related to the vector field ut via (2.1). A necessary condition for a vector field ut to be minimal is that the derivative of the cost functional E vanishes at ut, that is DE(ut) = 0. It is shown in [10, Theorem 2.1.] and [60, Theorem
4.1] that DE(ut) = 0 is equivalent to
1 −1 0 1 0 Lut = det Dϕ (J J ) J , (2.4) σ2 | t,1 | t − t ∇ t
−1 0 −1 1 −1 where ϕt,s = ϕt ϕ and J = I ϕ , J = I ϕ . This condition ◦ s t 0 ◦ t,0 t 1 ◦ t,1 is then used in [10] to devise a gradient descent algorithm for numerically
20 computing the optimal transformation ϕ1.
Geometric reformulation of LDM
Formula (2.4) can be reformulated in a way that emphasizes its geometric nature. As we will show in Section 2.1.2, formula (2.4) is equivalent to
1 [ Lut = (ϕt.I ) ϕt, . (ϕ .I I ) . (2.5) −σ2 0 1 1 0 − 1
This formula can be understood as follows: the first factor ϕt.I0 is the action of the transformation ϕt on the image I0 V = (Ω). This is defined as −∈1 F ∗ the composition of functions, ϕt.I = I ϕ . The flat-operator [ : V V 0 0 ◦ t → maps images in V to the objects in V ∗, which are dual to scalar functions. (These dual objects are the scalar densities.) To describe such an operator, one first needs to choose a convenient space V ∗ in non-degenerate duality with V . We choose to identify V ∗ with functions in (Ω), by using the F L2-pairing Z f, I := f(x)I(x) dx , h i Ω where dx is a fixed volume element on Ω. With this choice, the flat operator ( [ ) is simply the identity map on functions. The space V ∗ is the smooth dual of V , that is the subspace of the topological dual, which would be the space of distributions, generated by linear functionals of the form I R fI dx 7→ Ω with f V . Although we have the flat-operator [, which allows us to ∈ move between the spaces V and V ∗, it is important that we conceptually distinguish between elements in V and in its dual V ∗. Indeed, the action of a transformation ϕ on an element in V ∗ is the dual action, and does not coincide with the action on V in general. In our example, the action on f V ∗ is ∈
ϕ.f = det Dϕ−1 (f ϕ−1) . (2.6) | | ◦
To see how this action arises, we need the abstract definition of a dual action, which is ϕ.f, I = f, ϕ−1.I . h i h i
Remark. The inverse in the definition of the dual action is necessary to
21 ensure that we have a left action:
ϕ.(ψ.f) = (ϕ ψ).f . ◦
Using this definition and the change of variables formula we see that Z Z ϕ.f, I = f, ϕ−1.I = (I ϕ)f dx = I(f ϕ−1) det Dϕ−1 dx h i h i Ω ◦ Ω ◦ | | = det Dϕ−1 f ϕ−1 ,I . | | ◦ [ Therefore, in the second factor ϕt, . (ϕ .I I ) of equation (2.5), the term 1 1 0 − 1 (ϕ .I I )[ is interpreted as an element in V ∗. Consequently, the action 1 0 − 1 is the dual action given by
[ −1 0 1 ϕt, . (ϕ .I I ) = det Dϕ (J J ) . 1 1 0 − 1 | t,1 | t − t
It remains to explain the last ingredient; namely, the diamond map in equation (2.5), : V V ∗ ∗ . (2.7) × → H This is the cotangent-lift momentum map associated with the given repre- sentation of the Lie group G on the vector space V . Such momentum maps are familiar in geometric mechanics; see, e.g., [35] or [50]. The momentum map (2.7) takes elements of V V ∗, regarded as the cotangent bundle T ∗V × of the space of images, to objects in ∗, the dual to the space of vector H H fields. The map depends on the choice of ∗. For example, using the H L2-pairing with respect to the fixed volume element dx, the momentum map (2.7) is defined for images that are scalar functions I V = (Ω) and ∈ F densities f V ∗ = ∗(Ω) by the relation ∈ F Z I f , u = f I u dx , (2.8) h i Ω − ∇ · so that in this case I f = f I. − ∇ Remark (Momentum maps).
In geometric mechanics, momentum maps generalize the notions of • linear and angular momenta. For a mechanical system, whose config- uration space is a manifold M, which is acted on by a Lie group G,
22 the momentum map J : T ∗M g∗ assigns to each element of the → phase space T ∗M a generalized “momentum” in the dual g∗ of the Lie algebra g of the Lie group G. For example, the momentum map for spatial translations is the linear momentum and for rotations it is the angular momentum. The importance of the momentum map in geometric mechanics is due to Noether’s theorem. Noether’s theorem states that the generalized momentum J is a constant of motion for the system under considera- tion when its Hamiltonian is invariant under the action of G on T ∗M. This theorem enables one to turn symmetries of the Hamiltonian into conservation laws.
[Notation for momentum maps: J versus ] For convenience in refer- • ring to earlier work, e.g., [37, 39], we distinguish between the notation J for general momentum maps J : T ∗M g∗ and the notation → for the particular type of cotangent-lift momentum maps on linear spaces, : V V ∗ ∗ that typically appear in applications of × → H Euler-Poincar´etheory, as in equation (2.7).
Remark (Momentum of images). Momentum maps for images have been discussed previously. In particular, the momentum map for the EPDiff equa- tion of [36] produces an isomorphism between landmarks (and outlines) for images and singular soliton solutions of the EPDiff equation. This momen- tum map was shown in [38] to provide a complete parametrization of the landmarks by their canonical positions and momenta. A related interpreta- tion of momentum for images in computational anatomy was also discussed in [61].
We now explain in which sense expression (2.8) is a momentum map. Even though the cost functional (2.3) is not invariant under the action of the diffeomorphism group, one may still define the momentum map : V V ∗ ∗ via × → H I f, u = f, uI , h i h i as done in geometric mechanics, see [50] and [35]. The action u.I is defined as u.I := ∂t t ϕt.I for a curve ϕt such that ϕ (x) = x and ∂t t ϕt = | =0 0 | =0 u. This is the infinitesimal action corresponding to the action of Diff(Ω) on V . Although the -map does not provide a conserved quantity of the
23 dynamics, it nevertheless helps our intuition and gives us a way to structure the formulas. Let us apply this concept to the registration of scalar images I (Ω) ∈ F on the domain Ω. The infinitesimal action is given by
−1 u.I = ∂t (I ϕ ) = I u t=0 ◦ t −∇ · and thus the momentum map in this case is Z
I f, u H∗×H = f, I u V ∗×V = ( I u)f dx = f I, u H∗×H , h i h −∇ · i Ω − ∇ · h− ∇ i as stated in formula (2.8). The key is to reinterpret the L2-duality between the functions I u and f as the duality between the vector fields f I −∇ · − ∇ and u. Using formulas (2.8) and (2.6) in equation (2.5), we regain the stationarity condition (2.4). Remark. Writing the gradient of the cost functional (2.4) in the geometric form (2.5) has several advantages. For example, it allows us to generalize an algorithm that matches images as scalar functions, to cope with differ- ent data structures, such as densities, vector fields, tensor fields and others. Making this generalization allows one to see the underlying common geo- metrical framework in which we may unify the treatment of these various data structures. We can also keep the data structure fixed and vary the norm , and thereby alter our criteria of how we measure the distance k · k between two objects. In addition, the geometrical setting introduced here for image analysis allows us to vary not only the data structure, but also to change the group of transformations. We will explore this possibility and consider images registration using an iterated semidirect product of diffeomorphism groups, thus incorporating multiple scales into the registration framework.
2.1.2 Abstract Framework
Diffeomorphic image registration may be formulated abstractly as follows. Consider a vector space V of deformable objects on which an inner product , is defined, that allows us to measure distances between two such objects. h· ·i We can think of V as containing brain MRI images, an example frequently
24 encountered in computational anatomy, see e.g. [60]. The distance between two objects can be defined as I J 2 = I J, I J , which in the case k − k h − − i of images is the L2-distance Z I(x) J(x) 2 dx . Ω| − |
The second ingredient is a Lie group G of deformations, that acts on the space V of deformable objects from the left
(g, I) G V gI V. ∈ × 7→ ∈
In computational anatomy G usually is taken to be the group of diffeomor- phisms Diff(Ω) or variants of it. A diffeomorphism ϕ Diff(Ω) acts on ∈ images by push-forward; that is, by pull back with the inverse map,
−1 −1 ϕ.I := ϕ∗I = I ϕ or ϕ.I(x) = I(ϕ (x)) . ◦
Roughly speaking, this action corresponds to drawing the image I on a rubber canvas, then deforming the canvas by ϕ and watching the image being deformed along with the canvas. It is also the basis for the familiar Lagrangian representation of fluid dynamics as described in [37].
Given a curve t gt of transformations, we define the right-invariant 7→ velocity vector ut g as ∈ −1 ut = (∂tgt)gt . (2.9)
We obtain ut by taking the tangent vector of gt and right-translating it back to the tangent space at the identity TeG = g, which is the Lie algebra of G. Rewriting (2.9) as
∂tgt = utgt (2.10) and specifying initial conditions at some time t = s, we obtain an ordinary differential equation (ODE). If we start with velocity vectors ut, we can solve this ODE to reconstruct the curve gt. This corresponds to the construction of diffeomorphisms as flows of vector fields via the equation
∂tϕt = ut ϕt , ϕ (x) = x . ◦ 0
This idea was first introduced in image matching in [18]. Let us denote by
25 u gt,s the solution of the ODE (2.10) rewritten as
u u u ∂gt,s = utgt,s , gs,s = e ,
u with the initial condition that gt,s is the identity e at time t = s. Since the u u time t = 0 will play a special role, we denote gt := gt,0. Standard results for differential equations show the following properties
−1 −1 gt,sgs,r = gt,r , gt,s = gtgs , gt,s = gs,t , which we will use in our calculations.
Following the motivation discussed in Section 2.1.1 we define the abstract version of the cost functional (2.3) as
Z 1 1 u 2 E(ut) := `(ut) dt + 2 g1 I0 I1 V , (2.11) 0 2σ k − k where the function ` : g R is a Lagrangian measuring the kinetic energy → contained in ut and is the norm on V induced by the inner product k · k , . Note that formula (2.11) defines a matching problem for any data h · · i structure living in a vector space V and any group of deformations G acting on V . Although it was inspired by the concrete problem of diffeomorphically matching scalar-valued images, the cost function (2.11) no longer contains any reference to image matching.
All the results in this section are to be interpreted formally. They are helpful to show the common geometrical ideas in different matching prob- lems, but as we will see in Section 2.2, they cannot be directly applied to the problem of diffeomorphic image registration. Therefore we will now assume that the objects are sufficiently smooth, so all the operations can be carried out.
Next, we want to deduce formula (2.5) for the derivative of the energy in our abstract framework. In order to compute the derivative DE(ut) we u need to know how g1 behaves under variations δut of ut. This is answered by the following lemma, the proof of which is adapted from [84] and [10].
Lemma 2.1. Let u : R g, t u(t) be a smooth curve in g and ε uε → 7→ 7→
26 a smooth variation of this curve. Then
Z t u d uε u u u δgt,s := gt,s = gt,s Adgs,r δu(r) dr Tgt,s G. dε ε=0 s ∈
Proof. For all ε we have
d uε uε uε g = uε(t)g , g = e . dt t,s t,s s,s
Taking the ε-derivative of this equality yields the ODE d d uε u d uε gt,s = δu(t)gt,s + u(t) gt,s , dt dε ε=0 dε ε=0
u d uε and then, using the notation δgt,s := dε ε=0 gt,s, we compute
d − − − gu 1 δgu = gu 1 u(t)gu gu 1 δgu + dt t,s t,s − t,s t,s t,s t,s u −1 u u + gt,s δu(t)gt,s + u(t)δgt,s u u = gs,tδu(t)gt,s
u = Adgs,t δu(t) .
u Now we integrate both sides from s to t and multiply by gt,s from the left to get Z t u u u δgt,s = gt,s Adgs,r δu(r) dr , s as required.
Notation and definitions for cotangent lifts. Already knowing from
(2.5) how the first derivative DE(ut) of the cost functional is going to look, we want to establish the necessary notation before we proceed with the rest of the calculation.
The inner product on V provides a way to identify V with the smooth • dual V ∗, which is a subspace of the topological dual. To I V one ∈ associates the linear form I[ := I, V ∗. h · i ∈ Given an action G on V , we define the cotangent lift action of G on • π V ∗ via ∈ gπ, I = π, g−1I , for all I V. h i ∈
27 As mentioned earlier in remark 2.1.1, the inverse in this definition is necessary to make the dual action G V ∗ V ∗ into a left action. × → Finally we define the cotangent-lift momentum map : V V ∗ g∗ • × → via I π, u = π, uI , h i h i
where uI is the infinitesimal action of g on V defined by uI = ∂t t gtI | =0 for a curve gt with g = e and ∂t t gt = u. The use of the momentum 0 | =0 map was motivated in Remark 2.1.1.
Now we are ready to calculate the stationarity condition DE(ut) = 0.
Theorem 2.2. Given a smooth curve t ut g, we have 7→ ∈
δ` u u DE(ut) = 0 (t) = g I g π , (2.12) ⇐⇒ δu − t 0 t,1 or, equivalently
δ` 1 0 u 0 1[ DE(ut) = 0 (t) = J g J J , (2.13) ⇐⇒ δu −σ2 t t,1 1 − 1
0 1 where the quantities π, Jt , and Jt are defined as
1 π := (guI I )[ V ∗,J 0 = guI V,J 1 = gu I V. σ2 1 0 − 1 ∈ t t 0 ∈ t t,1 1 ∈
When G acts by isometries, the stationarity condition simplifies to
δ` 1 0 0 1[ DE(ut) = 0 (t) = J J J . ⇐⇒ δu −σ2 t t − t
0 The quantity Jt is the template object moved forward by gt until time t 1 and Jt is the target object moved backward in time from 1 to t.
Proof. Using the notation π := 1 (guI I )[ = 1 (J 0 J 1)[ V ∗, we may σ2 1 0 − 1 σ2 1 − 1 ∈ calculate
Z 1 1 u 2 DE(u), δu = δ `(u(t)) dt + 2 g1 I0 I1 V h i 0 2σ k − k Z 1 δ` d uε = (t), δu(t) dt + π, (g1 I0 I1) 0 δu dε ε=0 − Z 1 δ` u = (t), δu(t) dt + π, δg1 I0 0 δu h i
28 Z 1 Z 1 δ` u u = (t), δu(t) dt + π, g1 Adg0,s δu(s) ds I0 0 δu 0 Z 1 δ` D u −1 E u = (t), δu(t) dt + (g1 ) π, Adg0,t δu(t) I0 dt 0 δu Z 1 δ` D u −1 E u = (t), δu(t) + I0 (g1 ) π, Adg0,t δu(t) dt 0 δu Z 1 δ` ∗ u −1 = (t) + Ad u I0 (g ) π , δu(t) dt , g0,t 1 0 δu which must hold for all variations δu(t). Therefore,
δ` ∗ u −1 (t) = Adgu I0 (g1 ) π δu − 0,t = guI gu π − t 0 t,1 1 [ = J 0 gu J 0 J 1 . − σ2 t t,1 1 − 1
If G acts by isometries, then the group action commutes with the flat map and we obtain δ` 1 [ (t) = J 0 J 0 J 1 . δu −σ2 t t − t This concludes the proof.
This theorem tells us how to compute the gradient of the cost functional for any data structure and any group action. Just like the cost functional (2.11) itself, it is expressed entirely in geometric terms and contains no reference to particular examples such as images. This makes the theorem widely applicable.
δ` Remark. Although the momentum δu (t) at each time depends on I0 and δ` I1, it turns out that δu (t) obeys a dynamical equation that is independent of I0, I1. The equation in question is the Euler-Poincar´eequation on G. History and applications of the Euler-Poincar´eequation can be found in [37], [50] and [51] and its use in computational anatomy in more detail in [89].
δ` Lemma 2.3. The momentum δu (t) satisfies
d δ` δ` (t) = ad∗ (t) . (2.14) dt δu − ut δu
29 This is the Euler-Poincar´eequation on the Lie group G with Lagrangian ` : T G/G g R. ' →
Proof. Because the cotangent-lift momentum map is Ad∗-invariant [50] we obtain from theorem 2.2
δ` (t) = guI gu π δu − t 0 t,1 ∗ u −1 = Ad u 1 I0 (g1 ) π . − (gt )−
Differentiation of Ad∗ follows the rules
∗ ∗ ∗ ∂t Adgt η = Adgt ad 1 η , ∂gtgt− ∗ ∗ ∗ ∂t Ad 1 η = ad 1 Adgt η . gt− − ∂gtgt−
From this we see that
d δ` d ∗ u −1 (t) = Ad u 1 I0 (g1 ) π dt δu −dt (gt )− = ad∗ Ad∗ I (gu)−1 π ut gt 0 1 δ` = ad∗ (t) , − ut δu and hence the momentum satisfies the Euler-Poincar´eequation.
Remark (Dependence of I0,I1 on the initial momentum). It might seem counter-intuitive that the momentum evolves independently of the objects we are trying to match. However, the objects I0,I1 do influence the mo- δ` mentum δu (t) in a significant way. Solving the Euler-Poincar´eequations δ` requires that we know the initial momentum δu (0) and this initial momen- tum depends on I0, I1 through the formula
δ` (0) = I (gu)−1π . δu − 0 1
Alternatively, we might look at the problem from the viewpoint of the variational principle. Assume that `(u) = 1 u 2 is the squared length of a 2 | | vector for some inner product , on g. If we have found a vector field ut h· ·i
30 and group element g1, which minimize
Z 1 1 2 1 2 u dt + 2 g1I0 I1 V , 2 0 | | 2σ k − k then the vector field ut must also minimize
Z 1 u 2 dt , 0 | | among all vector fields uet whose flows get coincide with gt at time t = 1, i.e., ge1 = g1. But this means that ut must be the velocity vector field of a geodesic gt in G. Here we have implicitly endowed G with a right-invariant Riemannian metric induced by the inner product , on g. The Euler- h· ·i Poincar´eequation (2.14) is just the geodesic equation on the Lie group G with respect to this Riemannian metric.
2.2 Registration using Diff(Ω)
2.2.1 The Setting
In computational anatomy the group of deformations G is usually the group d of diffeomorphisms of some domain Ω R . Different types of data used in ⊂ computational anatomy, such as landmarks, scalar-valued images or vector fields, are deformed by diffeomorphisms via the mathematical operations of pull-back and push-forward. Intuitively this corresponds to embedding your data into the domain Ω, then deforming Ω by the diffeomorphism and observing how the data is deformed with it. From a numerical point of view an efficient way of constructing diffeomorphisms is as the flow of a vector field. Following [92] we will consider a certain class of vector spaces, called admissible vector spaces.
Definition 2.4. A Hilbert space , consisting of vector fields on the domain H 1 d Ω, is called admissible, if it is continuously embedded in C0 (Ω, R ), i.e. there exists a constant C > 0 such that
u ,∞ C u H . | |1 ≤ | |
1 d 1 Here C0 (Ω, R ) is the space of all C -vector fields on Ω that vanish on
31 the boundary ∂Ω and at infinity with the norm
d X i u 1,∞ := sup u(x) + u (x) . | | x∈ | | |∇ | Ω i=1
An admissible vector space falls into the class of reproducing kernel H Hilbert spaces.
d Definition 2.5. A Hilbert space , consisting of functions u :Ω R H → is called a reproducing kernel Hilbert space (RKHS), if for all x Ω and d a a ∈ p R the point-evaluation evx : R defined as evx(u) := a u(x) is a ∈ H → · continuous linear functional. In this case the relation
d u, K(., x)a = a u(x), u , a R h i · ∈ H ∈
d×d defines a function K :Ω Ω R , called the kernel of . × → H If we denote by L : ∗ the canonical isomorphism between a Hilbert H → H space and its dual, then we have the relation
−1 a K(y, x)a = L (evx)(y) .
In order for the RHKS to be admissible the kernel K has to satisfy the following properties:
K is twice continuously differentiable with bounded derivatives, i.e. • 2 d×d K C (Ω Ω, R ) and K 2,∞ < . ∈ × | | ∞ K vanishes on the boundary of Ω Ω, i.e. K(x, y) = 0 whenever • × x ∂Ω or y ∂Ω. ∈ ∈ Further exposition of the theory of RKHS can be found, e.g. in [4], [69].
Example. The Sobolev embedding theorem (see e.g. [2, Chapter 6]) states d that for Ω R there is an embedding ⊆ d Hk+m(Ω) , Ck(Ω), m > → 2 of the Sobolev space Hk+m(Ω) into the space of k-times continuously dif- ferentiable functions Ck(Ω). Therefore for m big enough, Hk+m(Ω) is an
32 admissible space. The corresponding kernel is the Green’s function of the operator L = Id + Pk+m ( 1)j∆j. j=1 −
We fix a RKHS with kernel K and let u L2([0, 1], ) be a time- H ∈ H dependent vector field. We consider the differential equation
∂tϕt = ut ϕt , ϕ (x) = x . (2.15) ◦ 0
Results from [92] tell us that this equation has a solution
ϕ C1([0, 1] Ω, Ω) , ∈ × defined for all t [0, 1] and for each t the map ϕt :Ω Ω is a diffeomor- ∈ → phism of Ω. For our matching purposes we will use the group GH consisting of all diffeomorphisms obtained in such way.
2 GH = ϕ : ϕt is a solution of (2.15) for some u L ([0, 1], ) (2.16) { 1 ∈ H }
Theorem 2.6 (from [92]). GH is a group.
2 Proof. Let ut, vt L ([0, 1], ) be two vector fields and ϕt, ψt GH their ∈ −1 H ∈ flows. To show that ϕ GH consider the vector field ut := u −t and 1 ∈ e − 1 denote by ϕt its flow. Then, since ϕt ϕ and ϕ −t are both integral curves e e ◦ 1 1 of uet, i.e.
∂tϕt ϕ = ut (ϕt ϕ ) , ∂tϕ −t = u −t ϕ −t , e ◦ 1 e ◦ e ◦ 1 1 − 1 ◦ 1
−1 we have ϕt ϕ = ϕ −t and evaluating at t = 1 gives ϕ = ϕ . Hence e ◦ 1 1 e1 1 1 ϕ GH. 1 ∈ To prove that ψ ϕ GH we define the vector field 1 ◦ 1 ∈ ( ut, if t 1 (u ? v)t := ≤ . vt− , if 1 < t 2 1 ≤
If ηt is the flow of (u ? v), we see that for t 1 we have ηt = ϕt while for ≤ 1 < t 2 the flow is given by η = ψt− ϕ . Thus ϕ = ψ ϕ . Now we ≤ 2 1 ◦ 1 2 1 ◦ 1 can rescale the vector field to fit into the time interval [0, 1] and have thus shown that ψ ϕ GH, which completes the proof. 1 ◦ 1 ∈
33 2.2.2 Matching Problems
Consider a general matching problem that seeks to minimize an energy of the form Z 1 1 2 1 E(u) = ut H dt + 2 U(ϕ1) , 2 0 | | 2σ where U : GH R is a functional containing the information about the → data structure. Examples of such functionals are
U(ϕ) = Pn ϕ(xi) yi 2 for landmark matching • i=1| − | U(ϕ) = I ϕ−1 I 2 for image-matching • k 0 ◦ − 1kL2 The energy is to be minimized over the space u L2([0, 1], ) of time de- ∈ H pendent vector fields. The existence of a minimizer is shown in the following theorem.
Theorem 2.7 (from [92]). Let U : GH R be bounded from below and → have the following property
n n n −1 If (ϕ )n∈ is a sequence in GH such that ϕ ϕ and (ϕ ) N → → −1 n ϕ uniformly on compact sets and Dϕ ∞ is bounded then | | U(ϕn) U(ϕ). → Then there exists a minimizer u L2([0, 1], ) such that e ∈ H
E(u) = inf E(u) . e u
n Proof. Let (u )n∈N be a minimizing sequence, i.e.
lim E(un) = inf E(u) . n→∞ u
n Then since U(ϕ) is bounded from below, (u )n∈N must be a bounded se- quence in L2([0, 1], ). Since bounded sets in Hilbert spaces are weakly H n compact, we can extract a subsequence, again denoted by (u )n∈N, that converges weakly to some ue. From
n n u , u u 2 u 2 h ei ≤ | |L |e|L
n we see by passing to the lim inf that u 2 lim infn→∞ u 2 . Concerning |e|L ≤ | |L U(ϕn) it is shown in [92] that weak convergence un u of the vector 1 → e
34 fields implies the uniform convergence of ϕn ϕ on compact sets. Since 1 → e1 (ϕn)−1 is the flow of the vector field un := un we also have the uniform t t − 1−t convergence of (ϕn)−1 ϕ−1 on compact sets. A generalized version of 1 → 1 Gr¨onwall’s lemma (see [92]) gives the estimate
R 1 n n 0 C00 |ut |1, dt Dϕ ∞ C e 0 ∞ | | ≤
n 2 and since (u )n∈N is L -bounded, we see from
Z 1 Z 1 n n n n ut 1,∞ dt C ut k dt = C u L1 C u L2 0 | | ≤ 0 | | | | ≤ | |
n n that Dϕ ∞ is bounded as well. Hence by assumption U(ϕ ) U(ϕ ). | | 1 → e1 Putting all pieces together we get
2 1 E(u) = u 2 + U(ϕ ) e |e|L 2σ2 e1 n 2 1 n n lim inf u 2 + lim U(ϕ ) = lim E(u ) ≤ n→∞ | |L 2σ2 n→∞ 1 n→∞ inf E(u) , ≤ u∈L2
Hence ue is a minimizer for E(u).
In order to apply this theorem we need to check in our examples that the functional U(ϕ) satisfies the required property. For landmark matching even point-wise convergence would be sufficient for the convergence of U(ϕn) to- wards U(ϕ). For image matching the required convergence is proven below.
Lemma 2.8. If I ,I L2(Ω), then U(ϕ) = I ϕ−1 I 2 satisfies the 0 1 ∈ k 0 ◦ − 1k property in Theorem 2.7.
n n −1 −1 Proof. Let (ϕ )n∈N be a sequence in GH such that (ϕ ) ϕ uniformly n → on compact sets and Dϕ ∞ is bounded. First note that | |
n −1 −1 n −1 −1 I0 (ϕ ) I1 2 I0 ϕ I1 2 f (ϕ ) I0 ϕ 2 . k ◦ − kL − k ◦ − kL ≤ k ◦ − ◦ kL
Next approximate I with a smooth function f C∞(Ω) with compact 0 ∈ c support, so that I f 2 < . Then k 0 − kL
n −1 −1 I (ϕ ) I ϕ 2 k 0 ◦ − 0 ◦ kL ≤
35 n −1 n −1 n −1 −1 I (ϕ ) f (ϕ ) 2 + f (ϕ ) f ϕ 2 + ≤ k 0 ◦ − ◦ kL k ◦ − ◦ kL −1 −1 + I ϕ I ϕ 2 k 0 ◦ − 0 ◦ kL ≤ 1 Z 2 p n 0 2 n −1 −1 2 det Dϕ (I0 f) L2 + f ∞ (ϕ ) (x) ϕ (x) dx + ≤ k | | − k Ω| | | − | p + det Dϕ (I f) 2 k | | 0 − kL ≤ 1 Z ! 2 0 n −1 −1 2 C1 + f ∞ (ϕ ) (x) ϕ (x) dx + C2 . ≤ | | supp(f)| − |
Since (ϕn)−1 converges uniformly (ϕn)−1 ϕ−1 on compact sets, we see → that U(ϕn) U(ϕ) < , provided n is large enough. This concludes the | − | proof.
2.2.3 Landmark Matching
The simplest kind of objects used in computational anatomy are landmarks. n i d Landmarks are labeled collections I = (x1, . . . , x ) of points x R . Given ∈ two sets (x1, . . . , xn), (y1, . . . , yn) of landmarks, the landmark matching problem consists of minimizing the energy
Z 1 n 1 2 1 X i i 2 E(u) = ut dt + ϕ (x ) y . (2.17) 2 | |H 2σ2 | 1 − | 0 i=1
d n Our space of deformable objects is V = (R ) with the usual inner product
n X I,J = xi yi , h i · i=1 for I = (x1, . . . , xn), J = (y1, . . . , yn). The action of the diffeomorphism group GH is by push-forward
ϕ.I := ϕ(x1), . . . , ϕ(xn) .
dn ∗ dn The corresponding cotangent-lift action on the dual space (R ) ∼= R is given by ϕ.J [ = Dϕ(x1)−1,T y1, . . . , Dϕ(xn)−1,T yn ,
36 and the calculation D E D E I J [, u = J [, uI H∗×H = (y1, . . . , yn), (u(x1), . . . , u(xn)) n X = yi u(xi) · i=1 * n + X i = y δxi , u i=1 H∗×H yields the diamond operator
n 1 n 1 n [ X i (x , . . . , x ) (y , . . . , y ) = y δ i , x i=1 R where δx is the delta-distribution defined by f(y)δx(y) dy = f(x) for a test function f(y). Note that since, is a RKHS, the delta-distribution i H∗ y δxi is an element of the dual space , since it is the evaluation functional i yi i H i y δ i = ev at the point x Ω in the direction y . x xi ∈ Applying Theorem 2.7 we see that (2.17) attains a minimum. The con- dition (2.13) that a minimizing vector field ut must satisfy is
n 1 X i −1,T i i Lut = Dϕt, (ϕ (x )) (ϕ (x ) y ) δ i . −σ2 1 1 1 − ϕt(x ) i=1
i Consequently, the momentum Lut is concentrated only on the points ϕt(x ). By using the Green’s function K(x, y) of the differential operator L, the minimizing condition above can be rewritten for the velocity ut as
n 1 X i i −1,T i i ut = K(x, ϕt(x )) Dϕt, (ϕ (x )) (ϕ (x ) y ) . −σ2 1 1 1 − i=1
2.2.4 Image Matching
The large deformation diffeomorphic matching framework used in [9] and
[10] seeks to match two images I0,I1 by minimizing
Z 1 1 2 1 −1 2 E(u) = ut k dt + 2 I0 ϕ1 I1 L2 . (2.18) 2 0 | | 2σ k ◦ − k
37 This example has already been discussed in Section 2.1.1. We review it here and apply the abstract formalism developed above. In this example the space V = (Ω) of deformable objects consists of real valued functions F on Ω. We endow this space with the L2-inner product. The group of deformations is again the group of diffeomorphisms Gk, generated by vector
fields in k. To avoid analytical difficulties we restrict ourselves to smooth, H i.e. C∞, images with compact support and require the vector fields to be smooth as well.
The action of GH on V is by push-forward
−1 ϕ.I = ϕ∗I = I ϕ ◦ for ϕ GH and I V . As we have seen, the dual action on the smooth ∈ ∈ dual reads ϕ.π = det Dϕ−1 π ϕ−1 , | | ◦ where det Dϕ denotes the absolute value of the determinant of Dϕ. The | | diamond operator in this example is
I π = π I. − ∇
It is proven in Theorem 2.7 and Lemma 2.8 that the energy (2.18) does admit a minimizing vector field. In order to apply theorem 2.2 one needs to assume more regularity of the image: I H1(Ω) would be enough. With 0 ∈ this additional regularity the necessary conditions are
1 −1 0 1 0 Lut = det Dϕ (J J ) J , (2.19) σ2 | t,1 | t − t ∇ t
0 −1 1 −1 where J = I ϕ , J = I ϕ , and ϕt,s is the flow of the vector field t 0 ◦ t,0 t 1 ◦ t,1 ut
∂tϕt,s = ut ϕt,s, ϕs,s(x) = x . ◦ See [86] for an extension of the LDM framework to functions of bounded variation. Equation (2.19) was used in [10] in devising a gradient descent scheme to computationally find the minimizing vector field.
38 2.2.5 Vector Field Matching
Diffusion tensor magnetic resonance imaging measures the anisotropic dif- fusion of water molecules in biological tissues, thus enabling us to quantify the structure of the tissue. The measurement at each voxel is a second order symmetric tensor. It was shown in [66] and [72] that the alignment of the principal eigenvector of this tensor tends to coincide with the fiber orientation in brain and heart. d The fiber orientation can be described by a vector field I :Ω R and → matching two vector fields can be formulated as minimizing the energy
Z 1 1 2 1 −1 2 E(u) = ut H dt + 2 Dϕ1 I0 ϕ1 I1 L2 . (2.20) 2 0 | | 2σ k ◦ ◦ − k
d In this example the space of deformable objects V = X(Ω, R ) consists of vector fields in Ω, the deformation group is the group of diffeomorphisms
GK , generated by vector fields in , and GH acts on V by push forward H
−1 ϕ.I = ϕ∗I = Dϕ I ϕ . ◦ ◦
Again to avoid analytical issues, we assume that the vector fields are C∞ with compact support and that the space also consists of smooth vector H fields. The smooth dual of V with respect to the L2-inner product can be identified with the space of one-forms V ∗ = Ω1(Ω). Concerning the existence of the minimizing vector field we have the fol- lowing theorem.
Theorem 2.9. If is an admissible vector space and additionally embedded H 2 d in C0 (Ω, R ), then the energy (2.20) admits a minimizing vector field.
Sketch of proof. It is shown in [92, Chapter 12] that given an embedding d n , C1(Ω, R ) weak convergence u u of time-dependent vector fields H → 0 → un L2([0, 1], ) implies uniform convergence of the flows ϕn ϕ on ∈ H 1 → 1 compact sets. Since the derivative Dϕt of the flow ϕt satisfies the ODE
∂tDϕt(x) = Dut(ϕt(x)).Dϕt(x), Dϕ0(x) = Id ,
d the additional assumption , C2(Ω, R ) allows us to conclude that we also H → 0 have the uniform convergence of the derivatives Dϕn Dϕ on compact 1 → 1
39 sets. In the same way as in lemma 2.8 we can show that the matching functional U(ϕ) = Dϕ I ϕ−1 I 2 has the convergence property k ◦ 0 ◦ − 1kL2 U(ϕn) U(ϕ), whenever we have the convergence of ϕn → → ϕ and (ϕn)−1 ϕ−1 uniformly on compact sets and also the → convergence of the derivatives Dϕn Dϕ uniformly on compact → sets.
Now, proceeding as in theorem 2.7 shows the existence of a minimum.
The infinitesimal action of u on I V is given by the negative of ∈ H ∈ the Jacobi-Lie bracket whose components are
∂ui ∂Ii (uI)i = Ij uj = [u, I]i . ∂xj − ∂xj −
The object dual to vector fields with respect to the L2-pairing are one-forms π V ∗ = Ω1(Ω). The diamond map is given by ∈
I π = £I π div(I)π , − − where £I π denotes the Lie derivative of the one-form π along the vector i ∂ i field I. In coordinates, writing I = I ∂xi and π = πidx , we can write the diamond map in the form
j j ∂I j ∂πi ∂I i I π = πj + I + πi dx . − ∂xi ∂xj ∂xj
Using these formulas, we can write the necessary condition for a vector field ut to minimize (2.20) as