Medical Images and Signals

Document Sample
Medical Images and Signals Powered By Docstoc
					                   Medical Images and Signals MIAS
           Interdisciplinary Research Collaboration –
                                      Final Report

Oxford University: Mike Brady 1 , Alison Noble, Steve Smith, Lionel Tarassenko

University College London 2 : Simon Arridge, Dave Delpy, Dave Hawkes 3 , Derek Hill,
Andrew Todd-Pokropek

Manchester University: Tim Cootes, Alan Jackson, Chris Taylor, Neil Thacker

Imperial College London: Jo Hajnal, Daniel Rueckert

Industrial Advisor: Pauline Hobday

Scientific Advisor: Peter Wells

Administrator: Jennet Batten

  IRC Director, 2001-3
  Hawkes and Hill, and several of their colleagues, were at KCL until 2005, then moved to UCL
  IRC Director, 2003-7

Glossary                                                                        iv
1.     Introduction                                                             1
       1.1     The opportunity and the vision                         1
       1.2     The formation of the IRC                                         2
       1.3     The concept of Grand Challenges                                  2
               1.3.1 Grand Challenge One – Structure and Function     3
               1.3.2 Grand Challenge Two – Intelligent Image
                        Acquisition                                             3
               1.3.3 Grand Challenge Three – Multiscale Modelling     4
       1.4     Integration of the Grand Challenges                    4
1.5    Engagement with the wider medical imaging and
       e-science communities in the UK and beyond                     5
2      Quality of research                                                      6      A
       2.1     Structure and Function                                 6
               2.1.1 Non-rigid registration                           6
               2.1.2 Shape modelling and analysis                     6
               2.1.3 Segmentation & feature detection                           7
               2.1.4 Modelling and registration                                 7
               2.1.5 Registration and segmentation                    8
               2.1.6 Modelling and segmentation                       8
2.1.7 Registration, segmentation and modelling                                  8
               2.1.8 Evaluation methods                                         9
               2.1.9 Applications                                               9
       2.2     Intelligent Image Acquisition                           9
               2.2.1 Under-sampled dynamic MRI                         9
               2.2.1 Motion corrupted data                            10
               2.2.3 Quantitative imaging                             11
       2.3     Multiscale Modelling                                   12
               2.3.1 Frequency-domain signal analysis                          13
2.3.2 Modelling hydrodynamic changes                                           13
               2.3.3 Fusing signals and images                                 13
               2.3.4 Preliminary clinical results                              14
       2.4     Contributions to UK e-science programme                14
               2.4.1 MIAS-Grid                                                 15
               2.4.2 eDiaMoND                                                  15
               2.4.3 MiAKT                                            15
               2.4.4 IXI                                                       15
               2.4.5 MIAS-EQUATOR                                              15
3      Impact of the work                                                      15      A
       3.1     Key achievements                                                15
3.2    Conference prizes and highly cited papers                               16
               3.2.2 Best paper prizes                                         16
               3.2.3 Best papers                                               16
3.3    Software freely distributed to the research community          16
4      New opportunities                                                       17      B
       4.1     Gearing, other funding, links with non-IRC
               research groups                                        18       E
5.     Training and Career Development                                         19      C
       5.1     Workshops                                                       19
       5.2     Contributions to International and National Meetings   19
       5.3     Summer Schools, Teaching Sessions and Courses                   20
6      Relevance and Exploitation: Clinical and Industrial Links      20       D
7    New Ways of Working and Multidisciplinarity        22   F
8    Management and Organisation                        23   G
9    Science and Society                                     24   H
10   Added Value, International Branding and National
     Leadership                                              25   I

AKT           Advanced Knowledge Technologies (IRC)

BBSRC Biotechnology and Biological Sciences Research Council

BCIO          Breast Cancer Imaging Ontology

BCS           British Computer Society

BIRN          Biomedical Informatics Research Network

BMVA British Machine Vision Association

BOLD          Blood Oxygenation Dependent (MRI pulse sequence)

BRAINCIRC     A biophysical/biochemical model of cerebral autoregulation and metabolism

CAD           Computer Aided Diagnosis

CAMINO        A software package for diffusion MRI reconstruction and probabilistic

CARS          Computer Assisted Radiology and Surgery

CBF           Cerebral Blood Flow

CMIC          Centre for Medical Image Computing (at UCL)

COMPLEX       Centre for Mathematics and Physics in the Life Sciences and Experimental
              Biology (at UCL)

CRUK          Cancer Research UK

CSF           CerebroSpinal Fluid

CT            Computed Tomography

CVMIA         Computer Vision approaches to Medical Image Analysis

DAME Distributed Aircraft Maintenance Environment

DICTA Digital Image Computing, Techniques and Applications

DTC           Doctoral Training Centre

DTI           Department of Trade and Industry or Diffusion Tensor Imaging

e-DiaMoND     e-Diagnostic Mammography National Database

ECCV          European Conference on Computer Vision

ECR           European Congress of Radiology
EEG            ElectroEncephaloGram

EPSRC Engineering and Physical Sciences Research Council

Equator        IRC addressing the technical, social and design issues in the development of
               new inter-relationships between the physical and digital

ERP            Event Related Potential

EU             European Union

FDA            Food and Drug Administration (USA)

fMRI           functional Magnetic Resonance Imaging

FoV            Field of View

FSL            A comprehensive library of neuro-image analysis tools from fMRIB at

GC             Grand Challenge

GIMI           General Infrastructure for Medical Informatics

GPRS           General Packet Radio Service

GS             Google Scholar

GT2 and
GT3            Globus Tookit

HBM            Human Brain Imaging

HTD            Health Technology Devices

IBIM           Integrated Brain Image Modelling

ICAPR International Conference on Advances in Pattern Recognition

ICT            Information and Communication Technologies

ICU            Intensive Care Unit

IEE            Institution of Electrical Engineers ( now Institution of Engineering and

IEEE           Institute of Electrical and Electronics Engineers

IEEE-EBMS      IEEE Engineering in Biology and Medicine Society

IEEE-ISBI      IEEE International Symposium on Biomedical Imaging

IIA            Intelligent Image Acquisition (Grand Challenge)

INRIA          l’Institut National de Recherche en Informatique et en Automatique
IPAM           Institute of Pure and Applied Mathematics

IPMI           Information Processing in Medical Imaging

IRC            Interdisciplinary Research Collaboration

ISAR           Inverse Synthetic Aperture Radar

ISBI           International Symposium on Biomedical Imaging

ISI            A Citation database

ISMRM          International Society for Magnetic Resonance in Medicine

ISOTT International Society on Oxygen Transport to Tissue

IT             Information Technology

ITK            Image registration and segmentation ToolKit

ITU            Intensive Therapy Unit

IXI            Information eXtraction from Images

JMRI           Journal of Magnetic Resonance Imaging

KCL            King’s College London

kt-BLAST       kt-Broad-use Linear Acquisition Speed-up Technique

kt-SENSE       kt-SENSitivity Encoding

LBNP           Lower-Body Negative Pressure

LNCS           Lecture Notes in Computer Science

MAP            Maximum A Posteriori

MARIBS         MAgnetic Resonance Imaging for Breast Screening

MDL            Minimum Description Length

MDM            MultiDisciplinary Meeting

MiAKT Medical Imaging Advanced Knowledge Technologies

MIAR           Medical Imaging Augmented Reality

MIAS           Medical Images And Signals

EQUATOR        Joint initiative of MIAS and Equator IRCs
MIAS-Grid      Separately funded project facilitating e-science and Grid aspects of MIAS

MICCAI         Medical Image Computing and Computer Assisted Intervention

MIT            Massachusetts Institute of Technology

MM             Multiscale Modelling (Grand Challenge)

MR             Magnetic Resonance

MRC            Medical Research Council

MRI            Magnetic Resonance Imaging

MSM            Multiscale Modelling (Grand Challenge)

MyGrid Separately funded bioinformatics project

NIHR           National Institute for Health Research

NIRS           Near-Infra-Red Spectroscopy

obel           object element

OWL            Ontology Web Language

PET            Positron Emission Tomography

PI             Principal Investigator

RDF            Resource Description Framework

RSNA           Radiological Society of North America

SF             Structure and Function (Grand Challenge)

S&F            Structure and Function (Grand Challenge)

SIP            Societal Issues Panel

SNM            Society of Nuclear Medicine

SPECT Single Photon Emission Computed Tomography

SPIE           International Society for Photooptical Engineering

SUV            Standard Uptake Value

TA             Triple Assessment

TINA           Open source image analysis environment (TINA Is No Acronym)

TOAST Temporal Optical Absorption and Scattering Tomography
TOI            Tissue Oxygenation Index

UCL            University College London

UCLA University of California at Los Angeles

US             Ultrasound

VIVE           Visual Imaging and Virtual Environment
1. Introduction
1.1 The Opportunity and the Vision
Over the last 5 decades medical imaging and acquisition of endogenous signals have
transformed practice in almost every branch of medicine. Ultrasound imaging (US) was
introduced in the 1960’s, three dimensional imaging first became available with the advent of
X-ray computed tomography (CT), positron emission tomography (PET) and single photon
emission tomography (SPECT) in the early 1970’s, followed a decade later by magnetic
resonance imaging (MRI). The following 20 years saw major advances in acquisition
technologies and in image reconstruction and formation. By the time the MIAS IRC started,
more and more images were being taken as the equipment became more widespread, even in
the smallest district general hospitals. Imaging systems, in particular MR and CT scanners,
were becoming available that could generate image volumes, each comprising hundreds of
Mbytes, many times per second. Multiple modalities were being used to acquire dynamic
sequences of images, for example to study function (such as contrast kinetics in tumours) and
motion (cardiac and respiratory). In addition, many areas of medicine required acquisition of
information from multiple imaging devices and other sensors. It was clear that the
conventional model practiced in most radiology departments, whereby raw images were
viewed as individual sets of slices or projections, was increasingly coming under strain. In the
words of the original proposal: clinicians were drowning in data when what they needed was
relevant and useful clinical information that would impact on clinical management and
decision making.

During the 1990’s, digital technologies associated with information processing, digital
manipulation, storage, transmission and visualization were advancing rapidly, but the
introduction of sophisticated information manipulation technologies into clinical practice
remained limited due primarily to poor robustness and lack of provision of complete solutions
that impacted in a positive way on clinical workflow and decision making.

By the end of the 1990’s, although it was clear that developments in image acquisition
technologies would continue (e.g. multi-coil imaging and fast acquisitions in MRI and multi-
slice and cone-beam CT), fundamental physical limits of signal-to-noise in MR and
ultrasound and radiation dose in CT were beginning to limit further advances. The bottleneck
to further progress was the inability to acquire and process clinical information effectively
using knowledge of anatomy, physiology and function. At the same time, biomedical research
was beginning to make use of imaging technologies to further fundamental understanding of
cellular and molecular processes in health and disease. Coupled with integrated information
processing, a systems approach to the understanding of biology and biological processes was

In the 1990’s significant results began to emerge from the world-class research undertaken in
the UK in a number of relevant areas of information processing in medical imaging and
signals. These included the work on statistical shape models at Manchester, infrared image
formation and reconstruction at UCL, image registration technologies at KCL and image
formation and signal processing at Oxford. However, the research basis in information
processing in the UK was fragmented and the synergies that were obvious to all were not fully
exploited. The EPSRC call for IT centric IRCs came at an opportune time in the development
of the discipline, research capacity and demand for results.

1.2 The formation of the IRC
The IRC was formed in January 2001 as one of five IT centric IRCs with a budget of £8.2
Million over 6 years. It was the only one of the five to be funded by two Research Councils
(EPSRC £5M and MRC £3M), emphasizing both its multidisciplinarity and strong application

The four original MIAS groups – KCL, Manchester University, Oxford University, and UCL
– reflected the network of scientific relationships and shared interests between the principal
investigators (PIs) at those four sites. Each had a proven record of close working with clinical
groups, a strong commitment to validating research results, an extensive record of research
collaboration with other groups, wide European and International links, and significant
collaborations with industry. Within six months of the start of MIAS, the Board of Directors,
taking note of the changes in personnel and organisation at Imperial College, resolved to
invite Jo Hajnal and Daniel Rueckert to bring Imperial in as a fifth member. Imperial has
subsequently played a key role in all of MIAS’s work, despite the fact that Imperial has not
been funded to the same extent as the original proposers. MIAS’s commitment to Imperial
may be gauged by the fact that the four founding groups agreed to top-slice their funding to
provide support for a postdoctoral researcher at Imperial and subsequently for a PhD
studentship. The KCL component of the IRC (Hawkes, Hill and several colleagues) moved to
UCL in January 2005 to form the Centre for Medical Image Computing (CMIC) with the
UCL members of the IRC (Todd-Pokropek, Alexander and Arridge). This further
strengthened and consolidated the critical mass of researchers in this area within London.

Originally, we proposed a project with a matrix of applications and methodologies, and set
out an elaborate plan of milestones based on relatively small projects, typical of a normal 3
year EPSRC grant:

Themes &          Stroke &       Heart       Cancer    Bones      High            Minimally
Clinical          Brain          disease               &          dependency      Invasive
Exemplars         disease                              Joints     care            Surgery
Robust extraction
of information
Closing the loop
acquisition and

1.3 The concept of Grand Challenges
During the first year, we recruited post-doctoral researchers and postgraduate students to
build the individual teams at each institution and instigated a number of “getting to know you
better” initial projects, as outlined in the proposal and covering much of the matrix. Towards
the end of the first year it became apparent that the matrix neither encapsulated the ambitious
vision and opportunity provided by the IRC nor did it constitute a framework to guide the
work in MIAS. In essence, it was too unwieldy a structure for a research programme in such a
large, diverse and complex area. At a 2 day workshop in Manchester, towards the end of year
one, we resolved to restructure MIAS as 3 Grand Challenges (GC). The IRC Board
considered that these reflected more effectively the external drivers outlined above, the
opportunities that these created, and the strengths of the MIAS team. The specific GCs are
outlined in the following three subsections. We have discovered that in our research
environment, and given the nature of our industrial and clinical end-user applications, the

scale of the GCs is a more appropriate unit of collaborative working. GC meetings would
typically be attended by 20 or 30 individuals, including industrial and clinical collaborators.
On the other hand, the whole IRC comprised a daunting 100 people at its peak, including
collaborators. Of course, the GCs were not intended to be silos – we continue to explore and
exploit interactions between them (Section 1.4), and with the possibilities afforded by e-
Science (section 1.5).

1.3.1 Grand Challenge One – Structure and Function
Segmentation (including feature detection), image registration and statistical
shape/appearance modelling are the three core technologies that form the basis of many
medical image analysis applications.

•   Segmentation: voxel intensities or textures are used to partition an image into different
    anatomical structures, regions or tissue types, providing a basis for quantitative analysis.
•   Registration: images are (non-rigidly) warped into a common co-ordinate frame,
    allowing information from different images to be combined and visualised, and shape and
    intensity differences to be analysed.
•   Shape and appearance modelling: the natural variability present in a class of images is
    modelled, providing biological insight and allowing realistic constraints to be applied
    during image interpretation. Of particular interest are statistical models of shape and
    appearance, which can be constructed from a set of images in which corresponding
    anatomical features have been identified.

Members of the consortium had made (and continue to make) significant contributions to
each of these technologies individually, but it quickly became clear that MIAS offered the
exciting possibility of developing a unified approach, combining the strengths of all three.

The key observation is that the three technologies are fundamentally complementary. Image
registration algorithms estimate correspondences between images, which can then be used to
construct statistical models of shape or to help obtain consistent segmentations across the set.
Both statistical models and segmentations can provide information which can, in turn, make
registration more efficient and/or reliable. The aim of the GC was to bring the expertise of
the different groups in each of these three areas together, to generate new, powerful
algorithms for extracting information from groups of images by simultaneously modelling,
registering and segmenting the images. Such methods would enable more detailed models of
organ structure and variation, and underpin a better understanding of the differences between
normality and a variety of disease states. The models would have the potential to lead to
better diagnosis of disease, and more accurate tracking of changes over time. It was also
recognised at an early stage that validation of such algorithms was crucial for comparing and
assessing different techniques, but represented a challenging and unsolved problem in it’s
own right, requiring a dedicated research effort within the project.

1.3.2 Grand Challenge Two - Intelligent Image Acquisition
In most current medical imaging, data are acquired with scanner settings that are fixed prior to
the acquisition process and control during data collection is limited to simple feedback
processes such as physiological triggering or “navigator pulses” in MRI. The resulting images
are generally analyzed in a separate, subsequent process. However, there is increasing
potential for real time control and for sophisticated computational tasks to be performed,
either in real time or in some other closely coupled manner. As a result, there is an emerging
capability to interact more fully with the scanning process, potentially to produce adaptive or
patient-specific imaging strategies and to make the subsequent data analysis more closely
coupled with the data acquisition process. For example, it may be possible to make use of
prior knowledge or models of motion, shape or function. The intelligent acquisition GC aimed
to explore research questions that arise out of this opportunity. Of course, stated like this, the

GC is far too broad, and so we have focussed effort onto imaging dynamic targets, either
organs in motion caused by cardiac pulsation and respiration, or as a result of contrast
agent/tracer, or a combination of the two. More specifically, the work of the Intelligent
Acquisition GC was divided into work packages undertaken by sub groups, usually from
more than one institution:

•   Under-sampled MRI: how substantially to reduce the amount of data acquired and yet
    reconstruct unambiguous images of moving objects such as the heart and of dynamic
    processes such as contrast uptake (UCL, KCL and Imperial).
•   Motion corrupted and hence inconsistent data: examples include PET, MR and CT
    (KCL, UCL, Imperial)
•   Quantitative imaging: design of image acquisition to directly measure or maximise the
    effectiveness of deriving physical or physiological quantities (Oxford, UCL and Imperial)

1.3.3 Grand Challenge Three – Multiscale Modelling
This GC responds to the increasing availability of imaging and sensing technologies that
provide information across a wide range of spatial and temporal scales, and was inspired by a
“systems” approach to image and signal analysis in medical applications. To concentrate the
effort, to build on the backgrounds of the researchers involved, and to relate to substantial
available clinical expertise, the GC has focused on one specific application: auto-regulation of
the cerebral vascular system. Neuroscience is a key research theme identified by all the
relevant Research Councils, as well as by the NHS, and there are many clinical problems to
which the results of this GC are relevant (e.g. diffuse brain disease; injury/stroke; atrophy and
degenerative disease; autonomic failure). This is a clinical domain in which a wide range of
physiological measurements are increasingly being made (e.g. perfusion acquisition and
analysis; functional brain imaging; ITU monitoring) and thus there is a clear need for model-
based signal analysis. Finally, this topic encompassed many of the generic technical problems
identified in the original IRC bid (e.g. automated co-registration; data acquisition;
segmentation and anatomical identification; detection and measurement of temporal change;
integration of near-infra-red spectroscopy (NIRS), magnetic resonance imaging (MRI), vital
sign monitoring and histology). The overall aims of the Multiscale Modelling Grand
Challenge have been:

•   The development/synthesis of physiologically-based, multiscale models of cerebral
    physiology (including haemodynamics, auto-regulation and metabolism) over a wide
    range of both spatial (cell to organ and whole system) and temporal (seconds to days)
•   Linking of the models to experimentally measured signals and images that can be
    obtained continuously/intermittently
•   The development of model-based, probabilistic signal analysis techniques to interpret
    and/or fuse multiple parameters extracted from signals and images
•   The use of the models to extract clinically useful information in diseases such as stroke

1.4 Integration of the Grand Challenges
In the final stages of the IRC we have begun to integrate the results from the individual GCs.
Examples of this integration include: (i) the cross-over of the modelling created using
registration and segmentation approaches to the generation of models of breathing motion in
PET and CT imaging of the lung for image directed radiotherapy102 and MR and CT imaging
of the heart96-8; (ii) the Multiscale Modelling work on auto-regulation has been integrated with
the balloon model of haemodynamic response function in functional MRI (S&F)127, (iii)
combining image registration with dynamic image acquisition to achieve high resolution 3D
fetal brain imaging83 and (iv) the use of image registration for reconstruction of cardiac
images from under-sampled dynamical MRI data144.

A notable success of the IRC has been by building relationships of sharing and trust in
collaborations between those working in different research institutions, each with its own
history and culture. While all members were used to interdisciplinary working with clinicians
and life scientists, in certain respects it is harder to set up collaborations with those labs with
somewhat similar and overlapping research expertise, especially as limitations on research
council funding tend to encourage competition rather than collaboration. We believe that it is
only by achieving Laboratories-without-walls, spanning multiple institutions and building
upon relationships of trust, that the UK will be able to build capacity and critical mass to be
world leading in this important area. It took about a year to establish this new way of
working; but, as a direct result of the IRC, there is substantial evidence that this will extend
well beyond the end of direct IRC funding. The large number of joint projects and joint
research grants already in the pipeline are testament to the success of this endeavour. This has
been particularly beneficial to postgraduate students and postdoctoral researchers in smaller
laboratories for whom intellectual isolation and lack of peer contact can be a real problem.

1.5 Engagement with the wider Medical Imaging and e-Science Communities in the UK
and beyond
Medical imaging and signal processing is an active area of research throughout the UK, which
boasts significant research activity beyond the funded laboratories making up the IRC. The
consortium has also made links with Martin Leach, Mike Brada and Steve Webb at the
Institute of Cancer Research and with associated grant funding (CRUK, HTD and EPSRC
project grants) this has been a productive collaboration on IRC related work in MR and X-ray
mammography, study of the vasculature of the liver and optimising radiotherapy of the liver
and lung. Early in the IRC programme, we presented a seminar on its work at the UKRC
annual meeting in Birmingham, and will present an updated presentation at the forthcoming
UKRC meeting in Manchester in June this year. Outside the UK, both the organisation and
the programme of the IRC have intrigued and impressed colleagues, who have sought ways to
collaborate. Specific examples include ongoing collaborations with Utrecht and Erasmus
Universities (Netherlands), INRIA (France), Tübingen (Germany), and with Yale, UNC, MIT,
Harvard, Massachusetts General Hospital in the USA.

The UK e-Science programme was announced shortly after the IRC was set up and its aims
aligned well with our objectives. E-Science is poised to make a major impact in medical
research and bioinformatics and the IRC was well positioned to contribute to this. The IRC
was directly involved in 5 e-science grid projects: MIAS-GRID – aimed to facilitate the e-
science and grid aspects of the IRC activity, primarily by building on the existing MyGrid
project; eDiamond – the development of a national federated mammogram database; MIAKT
(joint with the AKT IRC) – image and signal interpretation and the use of complex data in
decision support in triple assessment for breast cancer diagnosis and treatment; Grid Based
Medical Devices for Everyday Health (MIAS jointly with the Equator IRC); and IXI –
Information eXtraction from Images applied to dynamic atlas generation and its application to
degenerative brain disease. Lionel Tarassenko’s group also contributed substantially to the
DAME eScience project, which monitored the “health” of jet engines and created links
between the patient and engine health monitoring communities. The results of DAME,
MIAS-Equator and eDiamond form the basis of the follow-on project GIMI (Generic
Infrastructure for Medical Informatics). IRC members have also contributed substantially to
the EU HealthGrid programme, for example in the Mammogrid project. Mike Brady gave a
keynote addresses at the UK All-Hands eScience meeting and at the EU Healthgrid annual
meeting, the second of which was held in Oxford.

During the course of the IRC we have held a series of one day workshops to which we have
invited selected clinical colleagues and industrial partners. We participated in a major way in

all international meetings in our area, providing a significant number of key note lectures,
listed separately. Chris Taylor and Alison Noble chaired IPMI 2003.

2. Quality of Research
(This addresses evaluation criterion A. Here we summarise our research activity by Grand
Challenge. A full list of journal and conference papers can be found in Appendix A).

2.1 Structure and Function
The Structure and Function Grand Challenge aimed to build on existing strengths within the
IRC in image registration, shape & appearance modelling, and segmentation & feature
detection. As well as identifying important new challenges in each of these areas
individually, the GC set itself the ambitious goal of exploiting the potential synergies between
them, developing a unified approach to medical image interpretation and analysis. In the
sections below we document the main achievements of the GC, first identifying advances in
each area individually, then though pairwise combinations (eg simultaneous registration and
segmentation) and, finally, through full unification. We also describe new methods that were
developed for evaluating the results of applying these techniques, and discuss practical
applications of the work.

2.1.1 Non-Rigid Registration
Non-rigid registration, in its basic form, aims to find the deformation field which best warps
one image to match another. Typically this requires a representation of the deformation field,
an objective function to measure the quality of match, and a method of optimisation. The GC
has made significant contributions to all three aspects of the problem. Clamped-plate
splines331 were introduced as a useful method of interpolating deformation fields defined by
point correspondences, imposing clear boundary conditions. The importance of ensuring
diffeomorphic (smooth, invertible) deformations was explored extensively. Several practical
methods of modelling such transformations were propose30,42,261-2331,376,429, and the problem of
measuring geodesic distances between diffeomorphic deformation fields was addressed436.
Another focus of research has been in high-dimensional, topology-preserving methods of
manipulating deformation fields, based on a viscous fluid model. We have shown how, in
this framework, a ‘registration force’ field can be derived from a mutual information objective
function267. We have developed a fast multi-resolution fluid-based optimisation scheme,
using semi-coarsening and implicit techniques, integrated into a full multi-grid solution
framework, achieving a speed-up of several orders of magnitude without loss of accuracy42.
We have also placed a probabilistic interpretation on the commonly used mutual information
objective function, providing both analytic and empirical estimates of registration error238.

The practical importance of dealing with groups of images has emerged as a major theme for
the IRC. Although it is possible to non-rigidly register a group of images to a common
reference image using repeated pairwise registration, we have shown that a better approach is
to simultaneously register the group to a natural reference, computed by constraining the sum
of deformations over the group to zero223-4. The idea of groupwise registration is central to
our work on unifying registration and statistical modelling, and is discussed in more detail

2.1.2 Shape, Appearance and Biomechanical Modelling
An important focus of the GC has been on extracting information about the shapes and
appearance of structures, and the way they vary across populations. There has been a
particular emphasis on statistical models of shape and appearance. Construction of these
models in 3D relies on establishing a dense correspondence between the surfaces of
equivalent structures across a set of training images. We have developed an automated
approach to establishing these correspondences, based on modifying the points of
correspondence so as to minimise an information theoretic ‘description length’ of the set of

surfaces45-6,273. This minimum description length (MDL) approach has proved influential and
has been applied by ourselves and others in a range of practical applications (see below).
Over the course of the IRC we have reported a series of enhancements including a multiscale
iterative method of optimisation that scales approximately linearly with the size of the
training set270, optimal groupwise tessellation of the surfaces and the use of fluid-based
regularisation with shape images271. We now have a practical implementation that converges
for large data sets in minutes rather than days.

In parallel with this statistical modelling approach we have also developed biomechanical
models which aim to predict shape change due to external forces, particularly for use in the
breast153,176 and brain29,251. Early work in the IRC on the simulation of MR brain images
featuring dementia related brain atrophy led to a spin-off project with clinical collaborators at
the Dementia Research Centre, Institute of Neurology. This has resulted in a rigorous
framework for atrophy simulation in MRI using a biomechanical phenomenological model
and the development of a novel method for atrophy measurement29.

2.1.3 Segmentation & Feature Detection
Segmentation (classifying regions or pixels) and feature detection are important underlying
technologies for medical image interpretation. We have made advances in several important
areas, as well as developing particular applications using novel methods. In tissue
classification we have developed a Bayesian framework which accounts properly for partial
volume effects363. A new method of measuring cortical thickness from MR images has been
developed, which overcomes difficulties that arise due the variable sulcal arrangements which
affect voxel-based morphometry, and avoids the bias inherent in many surface-fitting
methods241. We have also developed a suite of level-set methods in the widely distributed ITK
package, and we have explored the relationship between level sets, Hidden Markov Random
fields, and graph cuts. In feature detection we have developed a robust probabilistic approach
to detecting focal intensity abnormalities, which has been applied to the detection of MS
lesions25,26, and have developed wavelet-based methods of texture analysis which have been
applied to breast and liver images . We have also made important contributions to the
emerging field of diffusion tractography – detecting and tracking white-matter tracts in
diffusion-weighted MR images of the brain – accelerated through collaborations promoted by
the IRC. Of particular importance is work on developing probabilistic measures of
connectivity between cortical and sub-cortical grey matter which deal sensibly with crossing

2.1.4 Modelling and Registration
There is a natural synergy between methods of modelling and registration, which we have
exploited. We have shown that groupwise non-rigid registration (see above) can be used to
find the dense correspondence across a group of images that is required to build statistical
models of shape and appearance, automating what can otherwise be a subjective and
extremely labour-intensive task152,223-4. We have also shown how a finite-element-model-
based estimate of deformation can be used as a starting estimate for non-rigid registration,
when there are significant deformations such as will occur in image-guided surgery of the
breast176,248-9,411. We have explored coupling the problems of registration and model building
directly, using an MDL measure of model complexity as the objective function for groupwise
non-rigid registration, with the aim of obtaining the registration that leads to the simplest
model. This generalisation of the idea of MDL shape correspondence to MDL image
correspondence poses a challenging problem, which we have addressed in a series of
publications setting out the theoretical basis and practical implementation of the
approach261,263,427,430,432-4, which is now suitable for practical applications.

In related efforts at integration a group of researchers, one each from Oxford, Imperial, and
Manchester, and two from UCL established the IBIM project for quantitative analysis of brain

structures from MRI images, with particular emphasis on developing statistical models of
population variation in brain morphology based on large databases of expertly labelled brain
data from collaborating partners including the Centre for Morphometric Analysis at MGH,

2.1.5 Registration and Segmentation
We have also exploited the natural synergy between registration and segmentation & feature
detection. We have shown how mutual information registration can be improved by including
local phase features, computed using the monogenic signal335-6. More fundamentally, we have
explored coupling registration and segmentation directly. It is well known that segmentation
of noisy images can benefit from superimposition of images, but this requires that they be
accurately registered. Conversely non-rigid registration is prone to fall into local minima and
so may benefit from additional information provided by segmentation. We have developed a
scheme in which we simultaneously compute maximum a posteriori (MAP) estimates of both
segmentation and registration (rigid or non-rigid)456-9.

2.1.6 Modelling and Segmentation
At the start of the IRC, Active Shape Models (ASMs) and Active Appearance Models
(AAMs) were already established as efficient and effective methods of segmenting images,
using statistical models learnt from a training set. There was, however, no practical AAM
implementation in 3D. Early in the project, we invested significant effort to develop a robust
and efficient 3D AAM implementation, suitable for practical applications. We have also
shown that AAM performance can be improved significantly by automatically tuning the
search model learnt from the training set to the current image, during segmentation259, and by
augmenting the statistical appearance model with features such as corners and edges388-9. We
have shown how a statistical deformation model generated from a database of CT scans can
be used to segment bone from ultrasound images and subsequently build accurate 3D models
for image guided orthopaedic surgery253.

2.1.7 Registration, Segmentation and Modelling
Building on the work described in sections 2.1.4 (Modelling and Registration) and 2.1.5
(Registration and Segmentation) we have developed an integrated framework for
simultaneous registration, model-building and segmentation.              By including tissue
classification algorithms in the group-wise correspondence and modelling framework
described in 2.1.4, we simultaneously segment a set of MR images of the brain into different
tissue classes, model the mean pattern of tissue distributions across the group, and learn a
dense correspondence field between all images362. A key idea is that the model represents the
pattern of tissue fractions, rather than the mean intensities, and that during optimisation we
estimate the mapping from tissue type to intensity for each individual image, as well as
incrementally improving our estimate of the model. Preliminary results suggest that this leads
to more reliable and accurate correspondences across the group.

We have also described a framework for coupling registration, segmentation and modelling in
image-guided intervention65. We have shown how non-rigid registration can be used to
generate motion models for segmentation propogation and organ tracking in radiotherapy

2.1.8 Evaluation Methods
In order to draw evidence-based conclusions from our work, we have invested significant
effort in evaluation and validation methods. We have developed a framework for evaluating
the registration of brain images – based on a set of MR images, together with detailed ‘ground
truth’ segmentations of the brain structures (obtained from Massachusetts General Hospital) –
and have used it to compare the performance of pairwise and groupwise registration
algorithms43,269. The Generalised Overlap Measures developed for this purpose have

potentially wide applicability. We have also addressed the important problem of evaluating
the performance of non-rigid registration methods when ‘ground truth’ segmentations are not
available. Our approach is based on using the registration to build a statistical appearance
model which is used to generate a large set of synthetic images. “Specificity”, the overlap
between the distributions of synthetic and real images, provides a measure of registration
accuracy which we have shown is comparable with the ground-truth-based overlap method381-
 . Biomechanical models of breast deformation in MRI have been developed initially to
validate breast MRI registration. These are now being used used in a spin-off DTI funded
project and will be more fully developed within the recently awarded EPSRC grant on
“Model Based Analysis of X-ray Mammograms” to develop a validation strategy for aligning
serial X-ray mammograms297,473.

2.1.9 Applications
Many of the techniques developed in the GC are generic and can be used in a range of
applications. Methods for analysing shape change, based on statistical modelling (alone and
in combination), have been applied to the hippocampus270 and lateral ventricles of the brain in
schizophrenia213,291, the brain in pre-term neonates224, knee cartilage in osteoarthritis445-54,
vertebrae in osteoporosis368-9, hip and femur253 and the prostate205. An integrated approach to
shape, biomechanical and motion models for image guided interventions has been proposed65
with applications in focal ablation of liver metastases22 and lung radiotherapy102 reported.
Other applications have been described in breast imaging226-8,411,461, liver and colorectal
cancer462-3, the foetal heart49,51-54 and studies of sexual function50,55. The techniques are also
being applied in novel ways to better understand basic biological questions in the life sciences
for example in electrophoresis gel analysis, the transduction of mechanical stimuli through
skin cells483 and studies of morphogenesis in embryology.

2.2 Intelligent Image Acquisition

2.2.1 Under-sampled dynamic MRI
Under-sampling of dynamic MRI results in aliasing in the images produced, and the resulting
ambiguity must be resolved if useful images are to be produced. In many circumstances the
dynamic portion of the imaged object only occupies part of the field of view (FoV). If the
MRI raw data is under-sampled in its native reciprocal (or k) space and time, t, the frame rate
can be accelerated to the desired time resolution. The acquired data can then be Fourier
transformed in k and t, to produce a signal distribution a space- temporal frequency (x-f)
space. These data are aliased in x-f space, but if the dynamic part is localised this aliasing
tends to mix dynamic content with more static locations and so bandwidth sharing strategies
can be adopted if the distribution of behaviours in x-f space can be determined either from the
data themselves or from additional calibration information. By exploring the properties of x-f
space (in collaboration with ETH in Zurich and within the core IRC programme) we
developed a method of estimating the dynamic content of a data set directly from the aliased
data and this allows more accurate reconstruction than with training data as well as increased
scan efficiency and the option to use the method in situations such as contrast enhanced
studies where it is difficult to obtain training data94,146. A related extension has been to
develop methods for free breathing cardiac imaging, in which under-sampling is used with an
integrated signal-based respiration monitoring approach to allow continuous acquisition
without breath holding105,183.

In cardiac imaging it is not always possible to acquire fully sampled data so more
sophisticated methods need to be considered. We have developed inverse problem techniques
such as subspace regularisation, Bayesian priors, or time-series methods, and have also
worked in the image domain directly to extract salient features from aliased data306-8,394-5. In
this way we have extracted quantitative measures without an explicit image reconstruction

step. The main application is to studies of free-breathing cardiac MRI and arrhythmia where
assumptions about periodicity may not be valid.

Reconstruction algorithms for dynamic data generally operate with fixed arrays of voxels
whose signal intensity variations capture the dynamic processes concerned. However, rapid
signal changes in fixed voxels can in fact arise from relatively slow motions if the change is
due to objects moving in the field of view. If movement of objects is the dominant change, it
may be more efficient to represent this motion directly76. This idea led to the development of
a new concept in medical imaging, the obel, or object element that may be defined on a
regular grid at some point in time, but which then moves around carrying the local tissue
properties with them. The movement of the obels may then be represented using a
parameterized basis, and we hypothesized that this formulation would allow time varying
behaviour to be represented more efficiently. We have investigated how to formulate a
dynamic problem in terms of obels and have explored the consequences of this for a
reconstruction of regularly under-sampled data using optimisation techniques to estimate
directly both the obels in a reference time frame and a sparsely parameterised representation
of their trajectories through time using splines. Having developed the approach we have
validated it and are now starting to compare its properties with conventional static voxel

Finally, we have developed a prototype catheter tracking system using MRI hardware loaned
by Philips Medical Systems, including Fluorine transmit/receive electronics and a fluorine
tuned coil. We used this for in-vitro tracking experiments that showed proof-of-concept87.

We have collaborated with clinical colleagues to explore the use of the above and related
methods in proof of concept and initial clinical cardiac studies104,108,110-1,116.

2.2.2 Motion corrupted data
Respiratory motion degrades the quality of PET images of the lungs. A simulation study
showed that breathing motion resulted in reduced sensitivity for the detection of small and
low contrast lesions, and reduced standard uptake values (SUV). Both these effects are highly
significant clinically. A technique to correct for respiratory motion has been developed
whereby the CT frames from a gated PET-CT acquisition are registered to produce a model of
the 3D motion and the resulting transformations applied to the complementary PET
component437-41. This technique dramatically improves the quality of the PET data. This
technique is now being evaluated in patients with lung cancer and funding to develop the
technique further and develop a practical implementation is being sought.

In a second exploration of motion corrupted data, we have studied MRI of the fetal brain.
Imperial and UCL have used MRI to study premature infants. However, a vital missing link in
these studies is detailed information about the brain under conditions of normal pregnancy.
The problem is that the foetus usually does not remain still during scanning, so the only
practical methods have involved snapshot imaging of slices. Using a combination of modified
snapshot imaging to freeze motion and image registration methods to achieve data
consistency in the presence of foetal motion we have developed a robust MRI method for
foetal brain imaging. The approach adopted is designed to improve both the signal to noise
ratio and the resolution as compared to conventional approaches and to allow foetal brain
images to be reformatted into any plane for viewing and quantitative analysis83,301.

More generally, motion is an enduring limitation in MRI. Most image acquisitions involve
obtaining only just sufficient data for image reconstruction so that subject motion during the
data collection generally results in motion artefacts that can damage image interpretation and
in severe cases render the images useless for either clinical use or scientific study. If array
coils are used but sufficient data are acquired for a conventional reconstruction or a modest

speed-up only is applied, then there is intrinsic data redundancy. We have explored the
potential of this redundancy to allow test of data consistency to be developed that can be used
both to detect motion problems and to correct them6,27. This has culminated in a general
framework for dealing with inconsistent data within the framework of multi-coil acquisition14.
The methods have been extended to deal not only with anatomical data, but also with the
important and highly vulnerable application of diffusion tensor imaging4. The core ideas have
been propagated to other related problems such as removal of fat from echo-planar images316
and reduction of noise in highly accelerated parallel imaging acquisitions88. Motion artefacts
in fMRI data severely limit the potential for individual (clinical) fMRI studies and the
development of some acquisition strategies. New methods for modelling and removing these
artefacts are needed in these areas as well as in standard group fMRI studies. Due to the
complicated interaction between motion-related artefacts, physiological signal changes (e.g.
BOLD effect) and other artefacts (e.g. B1-inhomogeneities) the impact on existing fMRI
analysis methods in unknown, because of the lack of realistic, controllable ground truth. We
have developed a MRI simulator to enable the validity and accuracy of existing and novel
analysis methods to be tested systematically by supplying such a controllable and realistic
ground truth58,162,278-81,341,419. In addition, the simulations are useful for testing and improving
acquisition strategies, for example, very-high-resolution structural imaging, neuronal current
imaging, and many techniques for imaging at very-high-field strengths (7T and over) which is
a growing and important area in MR research.

2.2.3 Quantitative imaging
The development of microbubble contrast agents, together with new ultrasound imaging
techniques, has enabled real time imaging of blood flow within the microvasculature and
estimation of tissue perfusion. Tissue perfusion measurement is a key physiological index
and is useful not only in the diagnosis of cardiac diseases and cancer; but also in the
monitoring of patients’ responses to a wide range of therapies. As well as providing new
imaging opportunities, ultrasound microbubble contrast agents have been shown to serve as
an effective means for delivering therapeutic agents. Imaging and quantification in this area
can provide an invaluable means for guiding drug delivery processes as well as for controlling
or even initiating local delivery dosage. Two major problems with current imaging and
quantification of ultrasound microbubble contrast agents are low sensitivity and the
generation of artefacts in the images due to the non-linear microbubble behaviour. We have
studied image generation using ultrasound and contrast agents through a series of
experimental measurements, physical modelling and numerical simulations32,171,257-8,287-8. The
causes of a number of imaging artefacts have been identified, and new techniques and tools
have been developed to correct them. In particular a framework for quantitative attenuation
determination has been developed172-3,286,406-9.

Measuring the elasticity of biological tissue or more specifically in-vivo estimation of patient-
specific biomechanical properties is useful both to aid clinicians in characterizing suspect
masses and to provide patient-specific parameters that can be utilized for more realistic
biomechanical modelling. This is typically done by applying a force/stimulus to soft tissue,
imaging the deformation, and subsequently recovering the deformation field and
biomechanical parameters by solving an Inverse Problem involving the measured
deformations. We have worked on two principal problems. Firstly, a novel assisted-freehand
device has been developed which allows a very precise compression to be delivered while
directly capturing a radio-frequency data stream that provides a reproducible input from
which strain images are produced. In conjunction with clinicians from the Oxford Breast Care
Clinic this device has been tested in a clinical environment in a study of over 120 patients,
comparing measurements derived from strain images acquired with the device, B-mode
measurements and pathology information (ground truth). Second, we have developed two
novel finite-element based methods of relative Young’s modulus estimation, one a variant of
the traditional displacement-based Inverse Problem solution and the second using an “all-in-

one” similarity based approach. This work has been tested on simulated data, gelatin
phantoms and a preliminary evaluation has been done on clinical data320,322,324.

2.3 Multiscale modelling (MSM)
We noted in the previous section that the MM GC has focussed its efforts on auto-regulation.
Blood circulation in the human brain responds in a complex way to a large number of stimuli.
Failures of the circulation are implicated in a number of major pathologies, including diffuse
brain disease, injury/stroke, atrophy and degenerative disease, and autonomic failure. Auto-
regulation is a concept which encapsulates some of this behaviour. It is usually defined as the
ability of the brain to maintain adequate blood flow despite variations in a number of
"external" factors such as blood pressure. Auto-regulation does not occur via some simple
goal-orientated feedback system rather, the brain circulation responds to a plethora of
different stimuli (both physical and chemical), on a variety of time scales, all using different
pathways. The characteristics of the measured signals in patients exhibit non-stationarity and,
in pathological cases, may not obey a linear input-output relationship. We have applied
frequency-domain signal analysis to investigate the time-varying nature of the relationship
between key variables, such as blood pressure and cerebral oxyhaemoglobin. Collaboration
between UCL, Oxford and Manchester has developed a clinically relevant model11,150 which
integrates the following:

    •   The biophysics of the circulatory system
    •   The feedback pathways by which a variety of stimuli affect cerebral blood flow
        (CBF) and cerebral metabolism
    •   The functioning of vascular smooth muscle

The model of auto-regulation developed in the MM GC (BRAINCIRC) has been able to
reproduce, at least qualitatively, the way cerebral circulation responds to a wide range of
stimuli including changes in systemic arterial pressure, blood oxygenation, blood CO2 levels,
and functional activity of the brain. The computational model includes not only the biophysics
of the circulatory system but also the feedback pathways affecting CBF, cerebral metabolism
and the functioning of vascular smooth muscle. We have compared outputs from the model
and experimental data using notably, for the first time in this context, NIRS, a technique
which provides information on cerebral tissue oxygenation and haemodynamics on a
continuous, non-invasive basis, including estimates of the concentrations of oxyhaemoglobin
[O2Hb] and deoxyhaemoglobin [HHb] and the absolute cerebral tissue oxygenation index
(TOI). Analysis of the parameters measured by NIRS and using the biophysics model, have
shown that TOI is auto-regulated over a greater range than CBF, remaining almost constant
over a ±40 % change in blood pressure. Starting from this, we have proposed that cerebral
auto-regulation should be redefined as the maintenance of constant brain oxygenation levels,
rather than of constant blood flow, since the latter is only one part of the process, which also
involves changes in the proportional fraction of arterial and venous blood volume92,485. This
example clearly illustrates the advantages of underpinning experimental work with rigorous
model development, an approach also adopted by the other two GCs.

Once particular model parameters have been chosen as the best candidates for fitting, various
fitting methodologies have been investigated to improve the agreement between model
predictions and an individual’s measured physiological data, thus potentially enabling the
model to become subject-specific126,477. The results of such fitting show that the model’s
predictive power can be greatly increased by adapting its parameters to individuals.

The processes that contribute to cerebral auto-regulation are many and complex.
Physiological models of these processes, whilst providing important understanding, are
computationally expensive and incorporate many pathways that are both poorly understood
and poorly characterised, making validation very difficult. Various mathematical tools have

been developed to examine what behaviour in sub-models is robust214,472. These tools enable
qualitative predictions to be made, based on the sub-model structures, helping to decide, for
example, whether particular sub-models admit more than one steady state, whether steady
states must necessarily be globally stable, and whether model structure determines
qualitatively how outputs respond to their inputs214,277. Such qualitative techniques are
particularly helpful in situations where the physiology is under active investigation, and the
lack of numerical data means that constructing numerical models is particularly hard. They
can provide confidence in model predictions which might otherwise be regarded as the results
of incorrectly chosen parameters values, or alternatively they might highlight situations in
which accurate knowledge of a parameter value is essential to ensure reasonable model
behaviour. The BRAINIRC model has been made available as an Open Source software
( and the extensive paper describing it was recently
awarded the biennial Bellman Prize as the best paper in “Mathematical Biosciences”11.

2.3.1 Frequency-domain signal analysis
To account for the time-varying nature of the relationship between cerebral oxyhaemoglobin
and mean arterial pressure, a method based on computing the cross-correlation of wavelet
coefficients has been developed to detect frequencies of maximum linear similarity150. In
addition, the low-frequency variability of the CBF signal in humans has been shown to be
related in part to CO2 variability and we have also shown how these low frequency signals
can be used to estimate cerebral venous saturation90,478. Multivariate coherence analysis (a
frequency domain measure of the linear relationship between two signals) has been applied to
data from healthy subjects, but it is likely that in pathological cases the stationarity of the
relationship is impaired. This is the basis of a novel method being developed for assessing the
status of cerebral auto-regulation using spontaneous oscillations that can be measured non-

2.3.2 Modelling hydrodynamic changes
We have also developed a modelling approach to address the hydrodynamic changes that
occur in the skull during the cardiac cycle. In normal healthy individuals systolic expansion of
the basal cerebral arteries produces a pressure wave within the subarachnoid CSF which
causes an outflow of CSF through the foramen magnum into the compliant spinal CSF space.
The pressure wave is also transmitted to the major dural venous sinuses by systolic expansion
of the arachnoid granulations. The effect is that the systolic pressure wave is dissipated into
the formation of CSF and venous pulsatility and largely bypasses the cerebral circulation. In
addition, elastic artery walls absorb part of the energy of the systolic pulse wave which is then
released during diastole, further flattening the arteriolar pressure profile to which the
intracerebral circulation is exposed. This combination of processes maintains a constant
perfusion pressure and flow in the cerebral capillary bed despite the major pressure changes
seen between systole and diastole.

The model describes the inter-relationships between arterial, capillary and venous blood flow
and movements of CSF between the cerebral ventricles, subarachnoid and spinal CSF
spaces86. The model has been tested using MRI flow data and fully described all the statistical
variation in the measurements from 24 young healthy control subjects. We have demonstrated
clear abnormalities in patients with late onset depressive illness112.

2.3.3 Fusing signals and images
fMRI provides maps of neuronal activity in the brain with good spatial resolution; the EEG is
a global, indirect measurement of this activity, but with excellent time resolution. Recent
work in the IRC has begun to develop an improved understanding of the “neurovascular
coupling” between changes in the underlying neuronal activity and changes in the fMRI
BOLD signal, from data acquired during simultaneous fMRI/EEG experiments127.

New methods of quantifying changes in neuronal activity from stimulus-evoked EEG data
have been developed, based on pole tracking in auto-regressive models, as well as windowed
Fourier and wavelet transform approaches. This approach enables the EEG to be assessed on
a single trial basis, in contrast to the standard “event related potential” (ERP) technique which
makes use of an averaged measure of the EEG. Work to date has focused on the analysis of
EEG habituation data, in which a short time interval is used between successive stimuli in
order to elicit a significant variation in the EEG response to each stimulus. Standard time-
domain analysis of these data has shown a “disassociation” between the amplitude of the
averaged ERP and the psychophysical pain rating by the subject. Spectral and time-frequency
analyses are being applied to investigate high-frequency stimulus evoked spectral changes (in
the gamma frequency band) which are more reflective of perception. This type of analysis has
not to date been applied to the fusion of EEG with pharmacological fMRI studies.

2.3.4 Preliminary clinical results
Primary autonomic failure encompasses a range of disease entities whose common feature is
severe postural hypotension74,283. In healthy subjects, the normal cardiovascular response to
head-up tilt consists of a rise in mean blood pressure, heart rate and peripheral vascular
resistance and a fall in cardiac output and stroke volume. In primary autonomic failure, head-
up tilting results in severe postural hypotension (low blood pressure) associated with a slight
rise in heart rate. An acute reduction in cerebral perfusion is the presumed cause of
symptoms in patients with postural orthostatic hypotension related to primary autonomic
failure and in these patients we have shown experimentally that this is associated with a
substantial fall in cerebral tissue oxygenation (TOI) as measured by NIRS74,166. This occurred
in every patient and was temporally related to the reduction in blood pressure during head-up

A multi-modal monitoring study has also been carried out in healthy volunteers combining
two methods of interrogating the brain, NIRS and transcranial Doppler ultrasound485. The
challenges were designed such that all the important mechanisms that control and alter
cerebral blood flow could be tested (and compared to the BRAINCIRC predictions). The
challenges were: hyperventilation, hypercapnia, isocapnic hypoxia (down to 80% arterial
saturation) and hyperoxia. Studies have also been carried out in head-injury patients using
hyperoxia and hypercapnia, but here with the addition of invasive measurements such as
intra-cranial pressure, brain pO2 and biochemical markers of metabolism such as glucose,
lactate and pyruvate. Methods are being considered to improve the suitability and training of
the multi-scale cerebral model when simulating pathologies such as head injury.

The comparison of the effects of stimuli on the cerebral circulation (tilt table, Valsalva
manoeuvre and Lower-Body Negative Pressure (LBNP) test) monitored with transcranial
Doppler and NIRS have been followed up by measurements in the MRI scanner180. In these
studies performed on young healthy volunteers, the effect of varying LBNP on the aortic and
cerebral circulations in various arteries has been measured using MRI. The NIRS
measurements have been correlated with the blood flow changes and with other physiological
data measured while the subject is in the MRI scanner. Simultaneous measurements of CSF
pulsatility have also been made. The results have indicated that the systemic circulation is
affected by the LBNP in a more complex manner than has been previously reported and that
in turn this variability affects the pressure and flow challenges to the cerebral circulation.
Preliminary analysis of the NIRS data suggests that these are similarly affected. These are
very technically challenging studies to undertake, the tuning of a LBNP test in an MRI
scanner is difficult and issues of subject welfare whilst in the bore of the magnet impose
limitations on the method. This is only the second known attempt to perform such a study.

2.4 Contributions to the UK e-Science Programme
Medical images and signals are intrinsically large datasets, often require prompt analysis, and
are subject to stringent legal and ethical requirements on patient confidentiality. For all of
these reasons, MIAS has been heavily involved in the e-Science Programme. It has
contributed primarily through a number of major projects:

    •   MIAS-Grid, which aimed to develop Grid-based methods for managing and sharing
        data and methods including: images, image metadata, interpretation/analysis
        algorithms, workflows, provenance, and validation methodology. The intention was
        to draw heavily on existing Grid developments – particularly the myGrid project – to
        provide much of the necessary infrastructure.
    •   eDiaMoND, which aimed to develop a system to support the construction of a
        distributed, ‘federated’ database of mammograms so that the database appears to a
        user as if it were all maintained in a single location; whereas in fact the mammograms
        are kept at the sites at which they were taken (and which have legal responsibility for
        their maintenance) and can be stored in accordance with the policies of the individual
    •   MIAKT, which was a collaboration between MIAS and the AKT (Advanced
        Knowledge Technologies) IRC, and which developed collaborative medical
        decision making and, in particular to support the Triple Assessment (TA) process and
        Multi-disciplinary Meeting (MDM) for management of symptomatic focal breast
    •   IXI, which explored the potential of Grid technologies for medical image research
        applications and the idea of dynamic atlas construction based on the collection of a
        large data set of brain scans from 600 normal subjects; some of these data is publicly
    •   MIAS-EQUATOR, aimed to explore how best to integrate mobile devices and sensors
        into a grid environment and how to exploit this integration to support medical
        monitoring for everyday health

Three follow-on projects resulted: GIMI (General Infrastructure for Medical Imaging) is
based on the results of the eDiaMoND and MIAS-EQUATOR projects; NeuroGrid builds on
IXI to explore e-Science enabled image analysis in the context of neuroscience research and
clinical trials; and PsyGrid similarly draws on the results of MIAS-Grid, also in neuroscience
research. More details of the individual Grid projects are provided in Appendix F.

3. Impact of the work
(This section also addresses criterion A).

3.1 Key Achievements
As will be evident from Section 2, and from the more detailed GC descriptions in the
Appendix, there have been a substantial number of achievements in the MIAS project –
primarily scientific, but also translation of basic science to emerging clinical practice and the
take-up of ideas emanating from MIAS by industry. Equally, we number as one of our key
achievements the formulation through the GCs of a programme of research, which, through
the GCs, has tackled – and made substantial progress on – a number of fundamental scientific,
clinical and practical problems. Our programme of research has set the international agenda
on issues that range from group-wise registration, simultaneous segmentation and registration,
under-sampled dynamic MRI, correction of motion-corrupted imaging, and quantitative
imaging to modelling auto-regulation in the brain. Our work on e-Science is widely cited. At
one level lower, we may cite the obel concept, the constraint of being a diffeomorphism for
non-rigid registration, the tissue oxygenation index, biomechanical models of breast
deformation, ultrasound strain imaging, correction of microbubble ultrasound images, as well

as novel segmentation and feature detection methods. The list of PhD theses produced and
postdoctoral RAs trained during the course of the IRC and the hundreds of papers that we
have published bear witness to the continuing stream of quality work that we have produced
as a result of working together for such an extended period. Prizes, awards, invited lectures
and keynote lectures given by members of the IRC are listed in Appendix C.

3.2 Conference Prizes and Highly Cited Papers
MIAS been well represented at the major international medical image analysis meetings such
as MICCAI, RSNA, IEEE ISBI, SPIE and IPMI, as well as ISMRM, HBM, and other more
specialised conferences. We have dozens of publications in the leading medical image
analysis scientific journals, including Medical Image Analysis, IEEE Transactions on Medical
Imaging, Medical Physics, Physics in Medicine and Biology, Radiology, JMRI, British
Journal of Radiology, European Journal of Radiology, as well as Lancet, Nature, Nature
Neuroscience, Neuroimage, Magnetic Resonance in Medicine, Proceedings of the National
Academy of Sciences, and others.

3.2.1 Best paper and thesis prizes
    • Murad Banaji (UCL) paper on cerebral modelling awarded the biennial Bellman Prize
        as the best paper in Mathematical Biosciences
    • Tim Carter (UCL) on modeling for image guided breast surgery, Student Prize at
    • Rodri Davies, Carole Twining et al (Manchester) on MDL surface correspondence,
        Best Paper Prize at ECCV 2002.
    • Rhodri Davies (Manchester), BCS Distinguished Dissertation Prize 2003 and BMVA
        Sullivan Best Thesis Prize 2002.
    • Carole Twining, Steve Marsland et al (Manchester) on diffeomorphic mapping,
        runner up for the Francois Erbsmann Prize at IPMI 2003
    • Catherine White (Oxford) Level set segmentation of PET images, selected as one of
        highlights of SNM, June 2005, Toronto
    • Hasan Habatay - Winner of Best Student Paper Prize at IEE International Control
        Conference, Glasgow, August 2006.
    • Illias Tachtsidis – Duane F. Bruley Award from International Society of Oxygen
        Transport to Tissue. ISOTT 2004 Conference, Bari, Italy

3.2.2 Best papers (with ISI citations where appropriate in brackets)
    • The Structure and Function Grand Challenge best papers include: Davies et al 200246
        (50); Rueckert et al 2003152 (29); Davies et al 2002273 (8); Crum et al 200339 (30);
        Schnabel et al 2003153 (18); Behrens et al 200319 (62); Behrens et al 200320 (109);
        Parker et al 2003121 (35); Parker et al 2002123 (40); Parker et al 2002124 (43); Razavi et
        al 2003 (48).
    • The Multiscale Modelling Grand Challenge best papers include: Payne 2005126,
        Banaji et al 200511; Payne and Tarassenko 2006131; Payne 2006127; Banaji 200610;
        Rowley et al 2007150; Kim et al 200786; Banaji and Baigent in press473; Thacker et al
    • The Intelligent Acquisition Grand Challenge best papers include: Atkinson et al 20046
        (7) ; Kozerke et al 200487 (5) ; Miquel et al 2004109 (8); Irarrazahal et al 200576 ;
        Batchelor et al 200514 (3) ; Larkman et al 200688; Tang and Eckersley 2006175; Malike
        et al 200694; Uribe et al 2007183; Prieto et al in press144; Jiang et al in press83.

3.3 Software freely distributed to the research community
A significant body of our work has been incorporated into packages that are freely available
on the web including:

    •   TOAST ( a finite element package originally
        designed for modelling light transport in highly scattering diffuse media but finding
        many more applications for example in biomechanical modelling
    •   CAMINO ( a suite of tools for diffusion
        MRI reconstruction and probabilistic tractography
    •   Components of FSL (, a comprehensive library of image
        analysis and statistical tools for FMRI, MRI and DTI brain imaging data.
    •   Components of TINA ( , an open source environment developed to
        accelerate the process of image analysis research, contributed by the Manchester
    •   Components of VXL Libraries for 2D and 3D computer vision (
        for 2D and 3D computer vision.
    •   (, which is a package of registration software
        called IRTK written by Daniel Rueckert and Julia Schnabel
    •   (, a registration package called vtkCISG from the KCL
    •   BRAINIRC ( a biophysical / biochemical model of
        cerebral autoregulation and metabolism.

Members of the consortium have also made significant contributions to the Insight Toolkit
( and in particular 2D-3D and 3D non-rigid registration code.

4. New opportunities
(This section addresses criterion B).

The size, concentration of expertise across a number of leading research laboratories, and the
6 year timescale of the IRC has enabled us to tackle larger issues, and to take substantially
more risks, than would have been possible with the usual (3 year) model of response mode
funding. In particular, the Grand Challenges have enabled us to define and develop three
major and essentially novel areas for medical imaging and signal processing research. In the
Structure and Function GC, we have shown the potential for combining modelling,
segmentation and registration as a powerful paradigm for advanced, automated image
analysis. In the Intelligent Image Acquisition GC, we have shown how sophisticated models
of motion, coupled with a deep understanding of the image formation process enables us to
devise new algorithms to image structures that are moving rapidly due to cardiac or
respiratory motion. Finally, in the Multiscale Modelling GC, we have developed a systems
level approach to biophysical cellular and molecular processes, which allows us to build
sophisticated models of biological function to, in our example, better understand cerebral
auto-regulation in health and disease.

We have identified specific new opportunities in systems biology, including the combination
of information obtained across very wide ranges of spatial and temporal scales. UCL has
recently been awarded an MRC Interdisciplinary Bridging Award to further studies in this
area. Oxford. The teaching hospitals associated with Imperial, KCL, Oxford and UCL have
recently been named by the Department of Health/MRC as comprehensive biomedical
research centres, following a stringent competition. Key projects in the Biomedical
Engineering theme of the Oxford centre, building on the work of the IRC, include data fusion
in the Emergency Department and the use of cerebrovascular models in the assessment of
stroke patients. The UCL bid similarly highlights imaging, image guided interventions and
neurological monitoring. The techniques that have been developed within the IRC, are poised
to have a major impact in the life sciences in such areas as the study of cellular mechano-
transduction and the study of cell migration in embryogenesis and morphogenesis.

Members of the IRC are currently active in the preparation of a number of EU Framework 7
programmes (FP7). The topics addressed by the IRC features strongly in the Healthcare and
ICT Calls of FP7.

The work on “intelligent image acquisition” has opened up a significant body of work in
which knowledge of anatomy, function, biomechanics and the image formation process can
be harnessed to improve the integrity of information derived from medical imaging. We are at
the very early stages of these developments and anticipate a major expansion on this topic in
the next few years. Technologies in image guided interventions are moving beyond the static
“road-map” provided by conventional interventional radiology and image guided surgery.
Models of anatomy, motion and deformation are being incorporated into these systems to
allow focal treatments in structures that move significantly (such as the heart, lung and liver),
while intra-operative imaging is being harnessed to provide real-time feedback. The
technologies developed within this consortium are now being developed to aid interventions
in orthopaedics, the prostate, liver, heart and lung.

Imaging biomarkers are set to have a major impact on drug development. Bringing a new
drug to market is becoming prohibitively expensive even for the largest pharmaceutical
companies. Imaging biomarkers are now becoming available in long term disease processes
such as dementia and in the study of therapeutic response in many cancers. The technologies
developed within the IRC are poised to have a significant impact on the accuracy, robustness
and precision of the measurement of disease progression. Imaging shows promise in detecting
early failure of a drug as well as reducing the size and or time lapse in many trials both of
which should lead to significant savings in the pharmaceutical and biotechnology industries
development pipeline. As detailed below a university spin-out, IXICO, has been formed to
address the demand that this will create. We foresee a significant expansion in this area in the
next 5-10 years.

4.1 Gearing, other funding, links with non-IRC research groups
(This addresses criterion E. More details are provided in appendix E).

The IRC has deliberately maintained an open-door to other collaborations. Each site has
brought other collaborators to our research meetings and we have encouraged joint projects,
joint grant applications and other outreach activities. It is important to realise that while the
funding provided to the four sites that submitted the original IRC proposal (KCL, Manchester,
Oxford, UCL) was generous (approximately £300K per annum per site), there was at each
site, and at Imperial, substantial other funding, from Research Councils, Charities, and from
Industry. We have deliberately not ring-fenced the IRC funding or personnel (postdoctoral R
As and PhD students), preferring instead to use the IRC funding to hire additional staff and
further to leverage existing relationships. In particular, the strength of our consortium has put
us in a very good position to develop research programmes based on the original IRC remit
and bringing in other research groups working in medical imaging, clinical researchers and
industrialists. We were, as noted above, very successful in the UK’s e-Science programme
with members of the IRC successfully bidding for 6 e-Science projects. In particular, the e-
Science Directorate actively encouraged the other software IRCs (AKT, DIRC, Equator) to
collaborate with MIAS because of the challenges posed to Grid technology by medical image
and signal analysis.

Two components of the IRC leverage have been the expansion of links with clinicians and
with industry. We noted in Section 1 that one of the key criteria for selecting the groups that
comprise the IRC was, and continues to be, close working relationships with clinicians
(including a commitment to translational research) and industry. Our working relationships
with clinicians have grown at rates that are more than proportional to the increases in our
funding envelopes. In like manner, as we noted above, the diversity, scale and scope of our

links with industry have grown considerably over the 6 years of the IRC. We have also
demonstrated our entrepreneurial skills by launching, or being closely involved in a number
of successful start-up companies. This point is spelled out in more detail below.

We have also been successful in translating our research activity in collaborative research
with industry. Members of the IRC were successful in four bids funded in the first phase of
the DTI Technology Programme’s Call on “Imaging and Sensors in Security and Healthcare
Applications”. We have also been successful in responsive mode grant applications to
EPSRC, MRC, and BBSRC. At Oxford and UCL, the IRC has contributed to the highly
successful Life Science Interface doctoral training programmes ( and, which in turn has contributed substantially to the development of
molecular imaging and modelling activities in the IRC.

5. Training and Career Development
(This addresses criterion C)

Over the course of the IRC we have trained 48 PhD students. So far 19 have submitted their
theses and 14 have been awarded their degrees. A further 10 are writing up and about to
submit. We have also provided postdoctoral training to 54 scientists. Details of the time that
these individuals spent with the IRC and their destinations are listed in Appendix B. The IRC
has provided a significant cohort of trained scientists and many have stayed within the UK
academic system. Together these individuals provide a critical mass of scientists who know
that they can work together and provide a strong foundation for future activity in this area.
They have also provided a core of expertise to UK industry in this area and in particular to the
start-up companies associated with the IRC.

5.1 Workshops
Throughout the course of the IRC we have held a series of one day workshops to which we
have invited selected clinical colleagues and industrial partners. There have been regular GC-
specific workshops whose primary aim have been to further the work of, and collaborations
between, GC personnel. These have been held approximately quarterly, one per GC. The GC
workshops have been particularly effective at fermenting and catalysing research ideas, and
though attendance by non-IRC researchers have provided both fresh ideas and insight and
enhanced external collaborations involving IRC members. These workshops typically
contained a mix of internal and external speakers and of pedagogical and research material. In
recognition of the highly diverse nature of the IRC membership, a significant aim of these
events has been to educate and establish effective communications. As well, we have held a
number of out-reach workshops, for example 3D imaging in head and neck cancer early in the
project, image registration (KCL 2002), PET image analysis (Manchester 2006), and a 2-day
workshop on “Vital Organ” modelling (March 2007) in which results and techniques from
brain modelling will be shared with researchers modelling other vital organ: heart, lungs and
liver. More clinically orientated workshops are planned in neurosciences, cardiovascular
disease and oncology in the final months of the project

5.2 Contributions to international and national meetings
Chris Taylor and Alison Noble chaired and co-chaired the prestigious and highly successful
international workshop “Information Processing in Medical Imaging” in 2003 in Ambleside,
Lake District and other members of the IRC were closely involved in it’s organisation470-1.
Similarly, Sue Astley and Chris Rose (Manchester), Mike Brady, and Reyer Zwiggelaer
(Aberystwyth) organised the International Workshop on Digital Mammography464, in
Manchester, June 2006

Several IRC members (Noble, Brady, Hawkes, Hill, Taylor) were founders of the UK national
meeting on Medical Image Understanding and Analysis and this close involvement continued

during the course of the IRC with highly successful annual meetings organised by IRC
members held at UCL (2000), Imperial (2004) and Manchester (2006). Imperial and UCL will
be hosting the MICCAI conference, the premier international meeting in medical image
analysis and image guided interventions, in 2009. Alison Noble was elected a Board Member
of the MICCAI Society in 2006.

5.3 Summer schools, teaching sessions and courses
As foreshadowed in our original proposal, we have held a highly successful series of one
week “summer” schools in medical image analysis, the first in 2004 at Imperial, the second in
2006 at Oxford and the third planned for this coming June in Manchester. The faculty for
each of these has included a couple of international speakers; but the majority of the lectures
have been given by members of the IRC. The students for these summer schools have come
from many European countries, and a few from North America and South East Asia.

We have convened special sessions at the UKRC, the major UK congress on medical imaging
most years that the IRC has run, have written a special issue for the British Journal of
Radiology, volume 77, 2004, edited by Alan Jackson. We have presented the IRC at
exhibitions at the European Congress of Radiology (ECR) in 2004 and have run a course in
medical image registration in 2004, 2005 and 2006. As noted above, Oxford established the
Life Science Interface Doctoral Training Centre in 2001 and this now accepts 20 DPhil
students per year (from over 250 applications) complementing the DTC at UCL (run by
COMPLEX). Responding to the demand for places on the DTC, the Engineering Science
Department at Oxford University established a 1-yr MSc in Biomedical Engineering in
October 2006 under the course directorship of Professor Alison Noble and is due to complete
a new £25-million Institute of Biomedical Engineering in October 2007. UCL established the
Complex DTC in systems biology, the VIVE MSc (Visual Imaging and Virtual
Environments) and will start a new MSc in Medical Image Computing in the Autumn of
2007, complementing its existing MScs in Radiation Physics and Biomedical Engineering and
Medical Instrumentation.

6. Relevance and Exploitation: Clinical and Industrial Links
(This addresses criteria D)

Each of the partners in the consortium has significant links with the clinical end users of our
technology and the industries that will exploit the results of our research. These links were
strengthened and expanded by the IRC. We see academic research, clinical end-users and
industry forming a triangle in which a close relationship with each group enhances
developments. For the academic partner feedback from clinical users is used to stimulate new
ways to solve problems, while close contact with industry provides access to the necessary
technologies and potential routes to widespread adoption of research outputs. A good example
of this working effectively is in the Intelligent Image Acquisition Grand Challenge where
access to an excellent cardiology team led by Reza Razavi and close cooperation with Philips
Medical Systems has resulted in rapid development of new technologies that are implemented
in the clinic and brought to market by the company. The collaborative work with Zurich on
under-sampled MR and in particular on kt-BLAST and kt-SENSE has been implemented into
a product by Philips Medical Systems. Patents have been filed for the XF-choice work and the
method to eliminate training data. There are ongoing discussions with Philips about
exploitation of this work. The methods have been tested on small scale clinical cohorts at
KCL (Professor Razavi) and Imperial (Dr Schmitz). The work on low frequency vasomotion
signals that came out of the Multiscale Modelling GC has led to a method of continuously
estimating cerebral venous saturation and this is being incorporated in software being
marketed with the next generation NIRS systems by Hamamatsu Photonics.

As noted above the IRC has been highly successful in establishing translational research
activities with Philips, DePuy, Kodak, GE, Siemens, VisionRT, IXICO and iMorphics via the
Department of Trade and Industry (DTI) “Technology Programme”, other DTI sponsored
schemes and the Department of Health “Health Technology Devices” Programme.

At the beginning of the IRC we formally set up a clinical advisory group (chaired by
Professor Alan Jackson) and an industrial advisory group (chaired by Pauline Hobday, at that
time CEO of Oxford Biosignals). While we did hold a couple of general meetings for our
clinical collaborators it rapidly became apparent that this was not a good model to encourage
collaboration. An expert in (say) cardiology may not be so interested in a detailed talk on
oncology and general discussions tended to be rather superficial. Moreover, clinicians have
huge and often urgent demands on their time, and so they are reluctant to give up a whole day
for a meeting in which they might be interested in only one or two presentations, especially if
the meeting involves several hours of travel. Responding to this feedback, we set up instead a
number of focussed sessions in areas of current clinical interest and these proved highly
successful. The most effective clinical collaborations were those maintained by individual
contact with an IRC member – usually located close to the clinical site. That person would
then act as a conduit to ideas and technologies available within the broader IRC. This
approach has been successful in many areas of the IRC. Examples include the cardiac
application with Prof Razavi (KCL), the link with the Royal Marsden Hospital in breast
cancer and lung radiotherapy, the link with David Edwards and John Wyatt in Paediatrics at
the Hammersmith and UCL, with Nick Fox and his team at the Dementia Research Centre
and Martin Smith in adult ICU, both at the Institute of Neurology, and with Professors Friend,
Mortensen, Kerr, and Harris in the Cancer Centre at the Churchill Hospital, Oxford.

The dynamic with our industrial partners was somewhat different but the lessons learnt were
similar. The most fruitful collaborations built upon strong personal links built up over several
years between individuals in industry and the IRC. For entirely understandable reasons of IP
and commercial confidentiality it was difficult to have detailed and in depth discussions with
more than one commercial partner at a time. We did hold several plenary meetings with
industrial partners and these were useful but did not lead to many significant new
collaborations. All our industrial partners are invited to our regular plenary meetings and we
continue information updates to our industrial partners by sending them the proceedings
abstracts from all of our plenary meetings. We held a joint event with the London Technology
Network on 20th October 2004 and this was very well attended (over 90 attendees) with
several new contacts and leads established. Further information on this is supplied in
Appendix D. This meeting, together with the ongoing contact through plenary sessions has
led to Pfizer Ltd wishing to conduct a review with MIAS to investigate possible collaboration
on image analysis. We participated in the highly successful Excellence in Engineering
Research event organised for the EPSRC international review of engineering in October
2004. We exhibited together with our industrial partners Philips, DePuy and AstraZeneca and
our exhibits were very well received. We established productive research relationships
between Philips and KCL, Manchester and Imperial; DePuy and KCL, UCL and Manchester;
AstraZeneca and Manchester; Siemens and Oxford (via Mirada) and UCL; and Oxford with
GE Healthcare and with Microsoft as well as productive relationships on specific projects
with SME’s and start-ups such as IXICO, Mirada, iMorphics, VisionRT and Dexela.

A particular highlight of the IRC has been the formation and development of several highly
successful start-ups including;

Mirada Solutions, founded by Mike Brady and Oxford colleagues in 1999, developed the
RevealMVS™ for the fusion of medical images (including both rigid and non-rigid
registration), such as MRI-MRI, MRI-PET, PET-CT. This has been installed in almost one
thousand hospitals world-wide. Subsequently, Mirada worked on a system to monitor the
progression of degenerative brain disease (Parkinson’s and Alzheimer’s), both of which are

commercial products. As Mirada Solutions’ work moved progressively towards molecular
image analysis, it was acquired in 2003 by CTI Molecular Imaging (NASDAQ quoted), and
CTI, in turn, was acquired by Siemens in 2005. The former Mirada currently has almost 70
employees in Oxford and is the advanced applications laboratory of Siemens Molecular
Oxford BioSignals was founded by Lionel Tarassenko in the summer of 2000 and focuses
primarily on providing intelligent clinical algorithms for patient monitoring. In October 2003,
Rolls-Royce invested £3.2 million in Oxford BioSignals to support the development of engine
health monitoring using the company’s patented technology. Since then, Oxford BioSignals
has developed multiple innovative clinical solutions by utilising advanced data interpretation
platforms such as neural networks, data fusion and pattern recognition. These innovative
clinical solutions can lead to early crisis warning for better hospital care (BioSign™), and
efficient cardiac safety assessment for pharmaceutical research and development
(BioQT™). These products are marketed primarily in the US, in partnerships with US
healthcare and pharmaceutical companies.
t+ Medical, formerly e-San, also founded by Lionel Tarassenko (in 2002), was the first
company to take advantage of the launch of GPRS mobile phones for medical applications. It
has developed and commercialised a clinically proven tool for the self-management of long-
term conditions such as diabetes and asthma using mobile phone-based technology. In 2006,
LifeScan (the diabetes Division of Johnson & Johnson) invested in e-San and the company
was renamed t+ Medical. The company’s t+ diabetes, asthma and blood pressure products are
being used for Disease Management by a number of Primary Care Trusts in England.
iMorphics ( was founded in 2002 by Chris Taylor, Tim Cootes and
colleagues in the University of Manchester, to exploit advanced 3D medical image analysis in
health and life science, with a particular focus on the needs of the pharmaceutical industry.
The company has formed strong partnerships with clinical research organisations,
pharmaceutical companies and medical device manufacturers, and its products are in use in
international clinical trials. Results from IRC research have been taken up by the company,
and will form the basis for its next generation of products and services. Major customer
interest is currently focussed on biomarkers for osteoarthritis and osteoporosis, and
applications in orthopaedics.
IXICO ( was founded in 2005 by Derek Hill, Dave Hawkes, Jo Hajnal and
Daniel Rueckert, and for which Mike Brady acts as an Independent Director. IXICO provides
imaging solutions to the pharmaceutical industry and has gained substantial contracts from the
pharmaceutical industry, including a Phase III trial. Their particular expertise is in
neurological image analysis, rheumatoid arthritis, oncology and lung disease. They make use
of IRC generated IP via option agreements with Imperial and UCL.

Hot off the press is the nascent formation of Afuson, by Alison Noble, Mike Kadour and
colleagues, and which is based on a novel assisted freehand ultrasound acquisition device
which is capable of high precision displacement estimation, together with associated software
to analyse the resulting strain fields. Dexela is another recent start-up ( of
which Mike Brady is Chairman and which develops image analysis and reconstruction for the
new technology of breast tomosynthesis. Yet another is VisionRT (, for
which Dave Hawkes and Daniel Rueckert act as scientific advisers, and which is developing
patient positioning and breathing motion tracking technologies for radiotherapy.

A significant amount of the IP generated by MIAS is channelled through the above spin out
companies, with IP collaboration agreements with University technology
companies/divisions. In addition, each participating University has technology transfer
programmes in place which actively review research for potential IP to offer to their own
industrial networks.

7. New ways of working and multidisciplinarity
(This section addresses criterion F).

The IRC has established new ways of working within our discipline. At the outset, there were
strong medical imaging and signal processing groups in the UK, but, as noted in Section 1,
there was little cooperation between them. Indeed, if anything, there was competition for
increasingly tight funding. The apparent synergies in their research agendas, overall critical
mass and training opportunities were simply not realized. Although it took a while (though an
encouragingly short period of time) to build relations of trust between the individual research
groups, this has now largely been achieved. A dynamic and productive research environment
has been established which encourages cooperation while preserving the individual focus and
culture of each group. We have shown that different research groups can collaborate
effectively. This is evidenced by the number of joint papers and successful joint grant
applications that are emerging from the consortium.

We contend that in our area of research (indeed, in most areas of science) there is an optimal
size of grouping for effective communication to take place and a unit of ~2-6 senior
researchers (PIs), with a similar number of PhD students and post docs seems to us to be
about right. Effective use of the full range of communication technologies (regular face-to-
face meetings, teleconferences, WIKI and other web based tools) provide good mechanisms
for generation and transfer of ideas and sharing of progress.

By its very nature medical imaging and signal analysis is multidisciplinary. The
multidisicplinary nature of our research occurs at many levels. Within the discipline we have
engineers, physicists, computer scientists and mathematicians working on algorithm and
system development and testing. Each group was intrinsically multidisciplinary in these
technical disciplines before the IRC started. The most exciting and productive
multidisciplinary interface is between these groups, the clinical disciplines, who are the
ultimate end-users of our technology, and the life-scientists investigating new translational
diagnostics and therapeutics. As stated above, we initially formed a clinical advisory group
across all clinical disciplines but this rapidly evolved into individual clinicians or clinical
groups around specific clinical areas of interest (dementia, neuro-development, cerebral auto-
regulation, breast cancer, orthopaedics, cardiovascular disease etc) with a close link to a
specific grand challenge. The clinicians working at this interface were tightly engaged and the
collaborations were very successful as evidenced by the joint publications.
Multidisciplinarity is furthered by our contacts with industry. The primary contacts with the
medical imaging industry (Philips, Siemens, GE etc) have been through their engineering
research divisions and clinical specialists groups, both primarily technical. We have also
developed close links with the pharmaceutical industry, most notably through the GSK
imaging centre at the Hammersmith, Amersham International, Astra-Zeneca and Pfizer. Our
links here are with both life scientists and engineers and our multidisciplinary approach has
been very beneficial in developing these relationships. Our spin out companies (Mirada
Solutions (now Siemens Molecular Imaging), Oxford Biosignals, e-San (now t+ Medical),
iMorphics, IXICO, all serve the pharmaceutical industry in the drug development and testing
pipeline. The technology of imaging biomarkers is developing fast and the partners and
affiliates of the IRC are playing a pivotal international role in these innovations (e.g. input to
the FDA’s “Biomarker Consortium”).

8. Management and organization
(This addresses criterion G).

The key to our success has been the consensus approach to IRC management and subsequent
resource allocation. The consortium was initially formed and the proposal written as an equal
collaboration between groups of approximately comparable size, each with an established

international reputation. Unique among the IRCs, we asked that the grants be awarded
separately to each institution, rather than having a single budget holder. Each institution has
therefore been responsible for the allocation of resources and accounting for their
expenditure. As foreshadowed in the proposal, a Board of Directors was constituted, again
with equal representation from each site, plus a clinical advisor and an industrial advisor.
When Imperial were invited to join the IRC, they were also invited to contribute two Board
members (Jo Hajnal and Daniel Rueckert). We also implemented a mechanism for
redirecting a proportion of the resource as the board saw fit. All Board members travel
widely, yet we wanted to have regular meetings, so we implemented a system of monthly
Board teleconferences to deal with medium term organization and management issues. As
well, we held a series of 6 monthly plenary meetings at which we held face-to-face Board
meetings. We also held annual general plenaries in which every individual involved with the
IRC would present, either orally or by a poster, their research work with the aim of getting
feedback. These generated lively discussions at all levels and in particular were beneficial to
the PhD students, as they found a ready-made community of fellow researchers. Booklets of 1
page abstracts were produced and archived for years 2004, 2005 and 2006 (see www.mias- Each individual grand challenge held their own teleconferences, also at monthly
intervals, and held a series of highly successful one or two day workshops. These were
detailed, focused and highly productive.

Our consensual management process has been effective and has resulted in some important
and successful changes. The first was the top-slicing of the budget to invite Imperial to
participate, the next was the formation of the 3 grand challenges. We also top-sliced some of
resource in order to finance the workshops and summer schools. Although, in our original
proposal, we reserved the right to redistribute resource between the consortium partners this
has not been needed as there has been sufficient flexibility within each site’s grant to
reallocate resource locally.

9. Science and society
(This addresses criterion H).

Each of the PIs on the project is committed to the wider communication of our research
results to the public and media. Of particular note are Derek Hill’s Partnership for Public
Engagement grants from the EPSRC “Science, Maths and ICT in Medical Imaging” (Sept to
Dec 2004) and the subsequent much larger grant “Medical imaging resources for the school
curriculum” (May 2006-April2008). David Hawkes gave a series of lectures to 6th formers
under the auspices of the IEE from 2001-2003. Mike Brady presented some of the work of
the IRC at the National Showcase Science Conference, held at Warwick University in 2005,
and has given numerous popular science lectures, as well as talks on entrepreneurship both in
the UK and in France. The NIRS team at UCL (together with collaborators at Essex
University) presented at both the Royal Society Summer Science Exhibition and the Scottish
Science Week in Glasgow in 2006. Alison Noble gave a lecture to sixth form high school
children in the IEE Engineering in Health Lecture Series, 2004. Dave Hawkes provided a
briefing to the Digital Economy Workshop for the EPSRC Societal Issues Panel (SIP).

10. Added value, international branding and national leadership
(This addresses criterion I).

EPSRC made it clear from the outset that one of the criteria that would be used to gauge IRC
success would be “added value” – which, the MIAS Board has interpreted as needing to
demonstrate that the whole is considerably greater than the sum of the parts. Evidently, this
would not have been the case had each institution simply taken the IRC money and continued
their own research track, independent of the other groups. A-priori, this was more of a risk
for MIAS because of our decision that institutions should be independent budget holders. The

Grand Challenges, the increasing interplay between the GCs, and the eScience work with
other IRCs, demonstrate clearly the added value of MIAS. Harder to quantify, but ultimately
more important, has been the development of collaborations between the IRC sites, resulting,
for example in joint publications and new research projects. The IBIM project, led by Tim
Cootes, Bill Crum, Mark Jenkinson, and Daniel Rueckert is a conspicuous example, as is the
iXi project led by Derek Hill and Jo Hajnal. Other examples include the funded collaboration
between Meng Xing Tang (Oxford, now Imperial) and Rob Eckersley (Imperial); between
Dave Hawkes and Mike Brady (with Dexela); and between Dave Hawkes, Mike Brady and
Chris Taylor (EPSRC project recently funded) and between Stephen Payne (Oxford) and
Dave Delpy (UCL) on an FP7 project (in preparation). As we reach the end of the IRC it is
clear that it has catalysed the creation of what is sure to become a growing number of projects
across institutions – Laboratories without Walls.

Generally, hubris is uncomfortable for the IRC members. Unlike some of our international
counterparts, we have not constructed a flashy website, complete with a pulsating, rotating
icon. However, colleagues in the leading laboratories around the world quickly became
aware of the large scale scientific and sociological experiment we were undertaking with
MIAS. Many have confided their astonishment that we could transform respectful but keen
competition into trust and collaboration. The senior management of INRIA, France, were
inspired to create “horizontal projects”, which cut across individual sites. The BIRN projects
in the USA took due note of the MIAS experiment, as manifested at major conferences such
as IEEE ISBI. At a more detailed level, the concept of simultaneous segmentation,
registration and shape analysis, first articulated in the Structure and Function GC, has become
accepted internationally as a major goal. Similarly, there has been an extremely lively
international debate on the topic of intelligent image acquisition and the imaging of moving
structures during the lifetime of the IRC and this has led to a development process that has
spawned substantial research output involving many research groups. The increased clarity
achieved in understanding the nature of dynamic imaging data and the requirements for a
reliable reconstruction framework have changed the nature of dynamic MRI. The work of the
IRC was a significant contributor to this process.

MIAS has always aimed to be a beacon of excellence in medical imaging and signals research
in the UK, but certainly not in any sense to be exclusive. It has sought, wherever it seemed
appropriate, to have an open-door policy, for example with the Institute of Cancer Research.
Nevertheless, the fact that (uniquely) our IRC has been funded jointly between EPSRC and
MRC has meant that MIAS personnel have been called on to contribute to setting national
agendas. Dave Hawkes and Lionel Tarrassenko are advising EPSRC on its Healthcare
Strategy in the proposed Information Directed Health initiative in its contribution to the UK
Government’s Comprehensive Spending Review. Chris Taylor, Andrew Todd-Pokropek and
Mike Brady have advised MRC on its eScience programme; Mike Brady has served as a
member of the MRC Neuroscience Panel and Milstein Fund Panel. He was a member of the
Royal Society Working Party on ICT in Healthcare, which reported in December 2006. Dave
Delpy and Lionel Tarassenko were members of the Royal Academy of Engineering/Academy
of Medical Sciences Working Party on Systems Biology, which reported in February 2007.
PIs of the IRC have regularly served as members and chairmen of the EPSRC Prioritisation
Panels in Healthcare and ICT.

All but one of the National Institute of Healthcare Research (NIHR) comprehensive
biomedical research centres are at one of the IRC sites. At the outset of the e-Science
programme, MIAS made the case, successfully, that medical image and signal analysis
constituted a major opportunity and applications challenge for e-Science. The result is the
portfolio of projects sketched in Section 2.4.

MIAS has enabled the creation of a web of collaborations between scientists in different
institutions that will outlive the IRC funding. We are looking to continue the workshops,

summer schools and plenary meetings in some form beyond the end of the project. The
consortium has produced a stream of world-class science and, perhaps more importantly,
produced some future research leaders. MIAS has worked effectively with clinicians and
with industry, and has spun off a number of companies whose futures currently appear bright.