Piecewise affine initialized spline-based
patient-specific registration of a high-resolution
ear model for surgical guidance
Michel A. Audette1, Rupert Brooks2, Robert Funnell3,
Gero Strauss4, and Tal Arbel2
1 ICCAS, University of Leipzig, Germany michel.audette@iccas.de
2 Center for Intelligent Machines, McGill University, Montreal, Canada
3 AudiLab, McGill University, Montreal, Canada
4 University Hospital Leipzig, Germany
Abstract. Image guidance of ear surgery would enable an ENT surgeon
to navigate about the components of the middle and inner ear, but the
elaboration of anatomical models for this application is limited by the
resolution of CT and its inability to distinguish among soft tissues. As a
result, it is impossible to identify manually some tissues in clinical data,
while visible tissues can only be identified with significant overhead.
We propose a method for producing patient-specific description of the
middle and inner ear on the basis of the minimally supervised registra-
tion of a high resolution model elaborated from micro-MR to patient
CT, where the transformation among the model and the patient data
is determined in a component-wise coarse-to-fine strategy. The first two
stages feature a rough alignment on the basis of a few homologous point
pairs, followed by a refinement based on a global affine transformation
determined by mutual information. The middle stage involves a piece-
wise affine registration where each local affine transformation is given the
global transformation as a starting point and is determined by mutual
information over an appropriate anatomical mask. The final registration
of each component is produced by mutual information-based thin-plate
splines, whose anchor points overlap the affine-transformed mask.
1 Introduction
Image guidance of ear surgery would enable an ENT surgeon to navigate about
the components of the middle and inner ear, and in particular avoid critical tis-
sues such as the facial nerve, but the elaboration of anatomical models for this
application is limited by the resolution of CT and its inability to distinguish
among soft tissues. As a result, the tissues that are visible can only be identified
with significant overhead, and the descriptiveness of the resulting models is lim-
ited by the relatively coarse voxel sampling, even with state-of-the-art clinical
CT, in relation to the scale of the components of the anatomy. Moreover, it is
impossible to identify manually some of the relevant tissues in routine clinical
data, such as the chorda tympani nerve.

(a) (b)
Fig. 1. Illustration of ear anatomy
and surgery: (a) outer, middle and
inner ear anatomy, reproduced with
permission from T.C. Hain [1];(b)
post-auricular incision.
(b)
(a) (c) (d)
Fig. 2. High resolution ear model derived from micro-
MR data: (a) model as it appears in interactive web-
site; micro-MR data: (b) raw data and (c) intensity
non-uniformity corrected data; (d) orthogonal planes
depiction of micro-MR data overlaid with ear model.
Ear surgery typically begins with a post-auricular incision, as shown in fig-
ure 1, which may lead to the repair of a tympanic membrane, the replacement of
ossicular bones by a prosthesis, the resection of a choleastoma, or a combination
of these interventions. A related intervention involves drilling the mastoid bone,
behind the ear, and resecting a choleastoma present in it.
The presence of pathology complicates the application of statistical shape
models [2] in patient-specific segmentation, as the notion of an average shape
is compromised by the random nature of tumour. Furthermore, existing regis-
tration methods in use with anatomical models, typically featuring a mutual
information similarity measure and global affine-initialized spline-based trans-
formation, do not apply readily to an anatomy that features many components,
ideally transforming independently from each other, in contrast with brain or
breast registration. To further complicate matters, some components, such as
ossicular bones, may in fact be missing, as a result of a previous operation.
In spite of the limited resolution, segmentations of clinical CT scans have been
used to visualize and model the ear. For example, Seeman et al. [3] demonstrated
manual segmentation of middle and inner ear structures from high resolution
CT, and suggested a combination of surface rendering of soft tissues and volume
rendering of bone. Most use of clinical scans for the ear has involved manual or
simple threshold-based segmentation, although Xianfen [4] used a combination
of manual and 3D level-set segmentation on CT data.
Non-clinical imaging modalities such as histology, micro-CT and high-field
micro-MR imaging permit much better identification of small middle and inner
ear structures and have been used by a number of groups for generic anatomical
modeling. Folowosele et al. [5] demonstrated the descriptive quality of high-field
magnetic resonance imaging in producing models of the middle and inner ear.
One of us published results [6] of expert segmentation of high-field micro-MR
data [7], and it is the model refined from the triangulated boundaries of this
segmentation that is applied clinically in this paper.