Construction and assessment of a 3-T MRI brain template.

Abstract

New MR imaging protocols enable visualization of brain structures. However, for dedicated clinical applications such as targeting deep brain stimulation (DBS), a more accurate localization requires the use of atlases. We developed a three-dimensional digitized mono-subject anatomical template of the human brain based on 3-T magnetic resonance images (MRI). By averaging 15 registered T1 image acquisitions, we have shown that the final image corresponds to an optimal image, limited by the performance of the 3-T MR machine. We compared different preprocessing workflows for template construction. With the optimal strategy, along with validated existing processing methods, one T1 template and one T1-T2 mixing template were created in order to improve visualization of spatially complex deep structures. Reduction of voxel size to 0.25 mm(3) was also advantageous to observe fine structures and white matter/gray matter intensity crossings. Results demonstrated that such a template also improved inter-patient registration for population comparison in DBS. These MR templates are made freely available to our community (http://www.vmip.org/mritemplate) to serve as a reference for neuroimage processing methods.

Full text

Page 1

1
Construction and assessment of a 3T MRI brain template

Florent Lalys
Ganaoui

INSERM, U746, Faculté de Médecine CS 34317, F-35043 Rennes Cedex, France
INRIA, VisAGeS Unité/Projet, F-35042 Rennes, France
University of Rennes I, CNRS, UMR 6074, IRISA, F-35042 Rennes, France
Department of Neurosurgery, Pontchaillou University Hospital, F-35043 Rennes, France
Department of Radiology, Pontchaillou University Hospital, F-35043 Rennes, France






















Corresponding author:

Florent Lalys (florent.lalys@irisa.fr)

Unite / Equipe U746 VISAGES
IRISA
Campus de Beaulieu
35042 Rennes CEDEX
France

Tel.: +33 (2) 23 23 38 29
Fax: +33 (2) 99 84 71 71



, Claire Haegelen
, Pierre Jannin
, Jean-Christophe Ferre, Omar El-

inserm-00546487, version 1 - 14 Dec 2010
Author manuscript, published in "NeuroImage 2010;49(1):345-54"
DOI : 10.1016/j.neuroimage.2009.08.007

Page 2

2
Keywords

Brain Template, basal ganglia, 3T MRI, Deep Brain Stimulation



Word count: 4055
Abstract word count: 164
Tables: 1
Figures: 8
References: 44




Abstract

New MR imaging protocols enable visualisation of brain structures. However, for dedicated clinical
applications such as targeting Deep Brain Stimulation (DBS), a more accurate localisation requires the
use of atlases. We developed a three-dimensional digitised mono-subject anatomical template of the
human brain based on 3T MR images. By averaging 15 registered T1 image acquisitions, we have
shown that the final image corresponds to an optimal image, limited by the performance of the 3T MR
machine. We compared different preprocessing workflows for template construction. With the optimal
strategy along with validated existing processing methods, one T1 template, one T2 template and one
T1-T2 mixing template were created in order to improve visualisation of spatially complex deep
structures. Reduction of voxel size to 0.25mm³ was also advantageous to observe fine structures and
White Matter/Gray Matter intensity crossings. Results demonstrated that such a template also improved
inter-patient registration for population comparison in DBS. These MR templates are made freely
available to our community (http://www.vmip.org/mritemplate) to serve as a reference for neuro-image
processing methods.








inserm-00546487, version 1 - 14 Dec 2010

Page 3

3
Introduction


The development of medical imaging equipment is driving increased demand for reference data sets.
Anatomical reference images are becoming of vital importance for comparison of results, to achieve
optimal spatial and intensity resolution, and to allow better identification of structures. One of them are
the templates, which are defined as anatomical models built from multiple volumes averaging. They
can be mono or multi-subjects and are generally based on a single modality. Atlases are derived from
templates but can be built from different modalities and are most often characterised by specific
structures labelling. Paper-based atlases were originally obtained from experts in anatomy who
manually drew and labelled reference images. For instance, atlases of the human brain turn out to be
very helpful for various procedures. Print atlases by Schaltenbrand and Wahren (1977), Talairach and
Tourmoux (1988), or Ono et al (1990) have been used with success in various computer aided decision
systems. Today, digital atlases are directly built from digital images and offer new capabilities and
applications. Nowinski et al (1997) used a combination of these three digitized print atlases to develop
a digital atlas. They are used as anatomy teaching tools, by providing interactive labelling of structures
and high images resolution and contrast. In image analysis, deformable atlases provide a powerful tool
for image segmentation, by exploiting constraints derived from the image data together with a priori
knowledge about structure parameters (Zhou and Rajapakse, 2005). Elastic registrations to a standard
atlas of the organ of interest are also very useful, enhancing objectivity of image interpretation (Friston
et al., 1995). Finally, they may serve as a common space for population comparisons studies, such as
Voxel-Based Morphometry (VBM), which allows identification of anatomical differences between
groups of subjects (Shen et al., 2005).

Digital templates can be classified according to the number of subjects used for the computation.
Multi-subject templates are mainly built from single acquisition of different control subjects reflecting
the population targeted by the clinical study (Seghers et al., 2004; Lee et al., 2005). Such multi-subject
templates are primarily intended to serve as anatomical references for spatial normalisation usually
required before studying human anatomical or functional variability. A prominent example is the
inserm-00546487, version 1 - 14 Dec 2010

Page 4

4
multimodal template used in SPM (Statistical Parametric Mapping, Institute of Neurology, University
College of London, UK; Evans et al., 1993). The construction of smoothed multiple subject templates
captures inter-subject variability, but their use for alignment could hinder the representation of targets
in a common space.

Single subject templates strive to attain optimal spatial resolution. Different strategies were followed
for defining such templates. The first strategy was the use of co-registered mono or multimodal
acquisitions of a single particular subject mainly based on MR acquisitions. The Colin27 MRI brain
template (Holmes et al., 1998) has been used in various neurosurgical applications. For instance, St-
Jean et al. (1998) created a deformable volumetric template of the basal ganglia and thalamus in
combination with the Schaltenbrand and Wahren atlas in order to estimate template-to-patient
transformations. It has also been used to create probabilistic functional templates, as the one by Finnis
et al (2003) for the combination of intraoperative data with MRI data, or to validate template warping
techniques (Chakravarty et al., 2008). The second strategy was the combination of images and
histological templates (Yelnik et al., 2007; Chakravarty et al., 2006). The interest of the first strategy is
that the whole clinical acquisition setup is applied to healthy, living control subjects and they prove to
be quite feasible in practice. Histological templates have advantages like higher structural and spatial
resolutions. However they turn out to be very complicated and lengthy to produce. Finally, templates
following the first strategy are based on the average of all registered volumes, which enhances the
quality of the final template by increasing Signal-to-Noise Ratio (SNR) and contrast (essentially
between Gray Matter (GM) and White Matter (WM)). These templates are constrained by both the
resolution and the quality of available imaging technologies.

Digital templates need to be evaluated and deeply validated as they are used for anatomical reference
images in various applications. It's crucial to assess the quality of the images by quantifying their
parameters, from the contrast resolution to the Signal-to-Noise Ratio (SNR). Many parameters step in
for the creation of a template, including the MR machine intrinsic settings, the number of scans to be
averaged, the pre-processing methods, the order of pre-processing, and the processing parameters. All
inserm-00546487, version 1 - 14 Dec 2010

Page 5

5
of these parameters have to be optimized in order to create the template with the optimal construction
strategy. The Colin27 MRI brain template has been lightly validated in the original work of Holmes et
al., (1998) by computing intensity profiles to demonstrate the improvement in image quality, but no
such studies have been performed in depth. Moreover, templates have to be validated in a clinical
context to assess its actual added value. In neurosurgery, they are widely used for surgical planning and
targeting, especially in DBS (Deep Brain Stimulation). DBS is a procedure for patients with movement
disorders (e.g., Parkinson's disease) for which medical therapy is not effective. It uses electrical
impulses to stimulate targets (often the SubTalamic Nucleus, STN) in the brain. For such neurosurgical
procedures, identification of basal ganglia on patient specific images isn't always possible due to the
lack of contrast between structures. The use of digital atlases has helped for addressing this problem as
deep brain structures are more visible and allow more accurate targeting.

We present in this paper how we built and assessed MRI templates using a one-subject average of
volumes acquired with off-the-shelf medical image protocols, and processed with up-to-date image
processing methods. The objective is to create one T1 template with an optimal construction strategy
that we will validate in this paper. This framework relies on the fact that the quality of the template
increases with the number of volumes averaged, as demonstrated by Holmes et al. (1998), on condition
that every volume is perfectly defined in the same common space. The quality reaches a maximum that
would theoretically correspond to an MRI without noise and intensity inhomogeneities. We chose to
use image processing methods that have already been validated in medical imaging context along with
fixed MR machine parameters. However, we evaluated the impact of the order of image pre-
processing, the number of scan required to reach an optimal image and the global evolution of the
image quality. We showed quantitative results using different complementary criteria demonstrating
the effectiveness of the best strategy. Then, we studied the impact of template choice in a nonlinear
registration task in a Deep Brain Stimulation (DBS) context. Finally, with the optimal strategy we
constructed a T2 template in order to create a multi-modal brain template by mixing T1 and T2-
weighted data for better visualisation of deep brain structures. The basic principle was to retain helpful
information from T2, i.e. deep brain structures (such as basal ganglia), and to merge this Region Of
inserm-00546487, version 1 - 14 Dec 2010