Deformable Ultrasound Registration without
Reconstruction
Rupert Brooks
1
, D. Louis Collins
2
, Xavier Morandi
3
,andTalArbel
1
1
McGill University, Centre for Intelligent Machines,
rupert.brooks@mail.mcgill.ca, arbel@cim.mcgill.ca
2
McGill University, Dept. of Biomedical Engineering,
louis@bic.mni.mcgill.ca
3
University Hospital of Rennes, Dept. of Neurosurgery,
xavier.morandi@chu-rennes.fr
Abstract. Ultrasound (US) imaging is often proposed as an interopera-
tive imaging modality. This use nearly always requires that the collected
data be registered to preoperative data of another modality. Existing
intensity-based registration approaches all begin by reconstructing a 3D
US volume from the collected 2D slices. We propose to directly regis-
ter the set of 2D slices to the preoperative images. We argue this has a
number of advantages, including the omission of the potentially complex
reconstruction step, greater adaptability of the similarity measures, and
easier parallelization. We describe a system for performing this task and
present results on phantom data that show that our slice based method
consistently outperforms a reconstruction based method in both speed
and accuracy.
1 Introduction
Ultrasound (US) imaging is used for a wide range of medical tasks ranging
from diagnostic imaging to intra-operative image guidance. Many of these ap-
plications require the fusion of the ultrasound image with additional imagery in
other modalities. As a motivating example, consider the use of 2D ultrasound
imaging for interoperative guidance in neurosurgery. Preoperatively, a careful
imaging study is often done and used for planning. During the procedure, how-
ever, the patient’s brain deforms somewhat from its preoperative position. The
US imagery can be used to track this deformation and update the preoperative
information, but to accomplish this, it is necessary to register the interopera-
tively acquired images to the preoperative images. Unfortunately, registration of
ultrasound data to other modalities has remained a challenging problem.
Previous approaches to the registration of 2D ultrasound data have as a first
step reconstructed a 3D volume from the 2D slice data. This approach has several
disadvantages. Volumetric reconstruction of US data is a challenging problem in
its own right (see [1] for a review). It is also independent of the registration
process, and so must be completed before the registration can begin. If problems
are found with the reconstruction, or new data is acquired, the reconstruction
must be reprocessed before the system can be updated.
D. Metaxas et al. (Eds.): MICCAI 2008, Part II, LNCS 5242, pp. 1023–1031, 2008.
c© Springer-Verlag Berlin Heidelberg 2008

1024 R. Brooks et al.
Instead, we propose to perform registration directly from a subset of the 2D ul-
trasound slices rather than reconstructing a volume. The most obvious advantage
is that one does not have to go about the reconstruction process. Reconstruction
takes some time and some data is inevitably lost in the resampling process. How-
ever, we believe the more important advantages of this approach arise from the
separability of the similarity measure for each slice. This allows the registration
process to be made more flexible in a number of ways. (1) The registration can
be defined incrementally allowing new slices to be added, or faulty slices to be
removed. (2) The similarity measures can be adjusted to the properties of the
US on a slice by slice basis. This is important because as discussed in [2] it is
difficult to find a reliable cost function for multimodal US registration. Treating
the slices individually in this way gains some of the advantages of block-based
local registration methods. (3) Working on a slice by slice basis also introduces
opportunities for parallel computation, in both shared memory and non-shared
memory contexts. The similarity measureforeachslicecanbeprocessedbya
separate core, or a separate process entirely. (4) Finally, it is possible to begin
registration before all the slices are collected, which allows data collection to pro-
ceed in parallel with the optimization. This could lead to significant advantages
in time critical environments.
Instead of a single, pairwise, registration between multimodal volumes, the
registration without reconstruction strategy proposed in this paper is a group-
wise registration problem expressed as the combination of N pairwise slice to vol-
ume registrations. We have extended existing nonlinear registration algorithms
to allow this groupwise registration and explore the performance of this approach
on US and T1-MRI images of a deformable phantom object. We find that it is
feasible and results outperform a volume reconstruction based approach.
In the following section we describe the proposed approach in more detail, and
compare it to existing work. Section 3 describes our implementation which builds
upon existing well tested deformable registration procedures. Our experimental
data, procedure and results are described in Section 4.
2 Methodology
Parametric, intensity based, approaches to image registration work by defining
a measure of similarity, S(I
1
,I
2
), between the images to be registered and spec-
ifying a parameterized space of transformations to explore [3]. In the pairwise
expression of the problem, one image, called the moving image, is warped by this
transform, which makes it a function of the transformation parameters. The sim-
ilarity measure between the two images is then also a function of the parameters.
As the optimal registration is considered to be the point of maximal similarity,
the registration problem can be expressed as an optimization problem:
φ
opt
=argmax
φ
(S(I
f
(x),I
m
(x
′
(φ))) − λ
r
R(φ)) (1)
Here, I
f
(x)andI
m
(x) are the fixed and moving images considered as functions
over some space. x is a set of pixel positions in that space, and x
′
(φ)isthisset