Mapping, registration and atlases

1
Mapping, registration and atlases

• IMAGE workshop breakout group

2
Members of group

• Derek Hill, London
• Richard Baldock, Edinburgh
• Dave Berry, Edinburgh
• Tilak Ratnanather, Baltimore
• Xavier Pennec, Nice
• Denise Ecklund, Edinburgh
• Vincent Breton, Clermont Ferrand

3
Aims of breakouts

• To generate written report about issues
• Make recommendations for future work
• Highlight areas that might benefit from future
  research funding
• Make links with industrial activity
• Facilitate international collaboration

4
Breakout group sub-headings

• How do we make image registration grid services
  intraoperable?
• Can the grid provide a technology push?
• How should we represent mappings?
• Should we use grid-services for a major cross
  validation of algorithms?
• How can, or should,atlases be shared?
• How could these services be used commercially
  (eg for drug discovery)
• What user interaction is required?
• Statistics of mappings

5

• How do we make image registration grid services
  intraoperable?
• Do we need to devise an abstract model for these
  services?
• Eg what is a spatial transformation
• ITK is going some of the way in this direction
  and illustrates some of the difficulties

6

• Can the grid provide a technology push
• New techniques that simultaneously analysis
  cohorts
• Treating time better
• User interaction

7

• How should we represent mappings?
• Definition of coordinate systems
• Representing diffeomorphisms and other types of
  mappings
• Do we need an ontology?

8

• Should we use grid-services for a major cross
  validation of algorithms?
• Compare with 1995 retrospective registration
  evaluation project which used ftp for
  international multicentre collaboration
• Do you need a gold standard to do this?
• Could we share the necessary data
  internationally?
• Two approaches
• Test algorithms against standard data including
  sensitivity to data perturbations
• Test data against algorithms
• Who would pay for CPU cycles?
• Compare results of new algorithms with those of
  established algorithms eg SPM

9

• How can, or should, atlases be shared?
• Component data for atlas
• Algorithms that form part of atlas
• We need to carefully define what we mean by an
  atlas
• Is there an agreed definition for a static atlas?
• Is it an evolving representation of the state of
  knowledge

10

• How could these services be used commercially
  (eg for drug discovery)
• Regulatory approval?
• The grid could provide access to established
  validated algorithms
• Mechanisms for licensing s/w on a per-use basis.

11
User interface

• A user interface will be needed to interact with
  the algorithm
• Set parameters, use best practice parameters
• Customize to application
• Interact with the registration process, eg to
  add prior knowledge about corresponding features
• Tutorials
• Visualization of results how do we visualize
  mappings (application specific)

12

• Mapping statistics
• Statistics on deformations might be part of an
  atlas

13
How do we overlap with other groups

• Visualization and data exploration
• We need visualization tools for interaction and
  viewing results/metrics
• Image analysis, measurement and query
• How do we use queries to reason with mappings
  generated by regn. algorithms?
• Metrics from mappings
• Metrics from registered images
• Data management, metadata, ontologies, provenance
• We obviously need all of this stuff!

14
Issue of generalization

• Almost all headings in registration could be
  applied to other image analysis techniques eg
  segmentations

15
Future work of this group

• Select sub-headings for detailed discussions
  leading to recommendations
• What is easy to do, not easy to do, or dependent
  on other activity?
• Can we identify the state-of-the-art in these
  areas?
• Can we facilitate international collaboration?
• Link our aspirations to the current state of grid
  technology

16
A registration test bed?

• Registration is a core technology in medical
  image analysis
• There is no one algorithm that solves all
  problems
• There continues to be substantial algorithm
  development
• Registration needs to fit into many different
  image analysis workflows

17
Scenario 1

• Prototyping a new application
• New sort of data
• New question
• Want to find the right algorithm
• Grid can be used to do large scale comparison of
  different algorithms

18
Scenario 2

• Production image registration service
• Provides easy use for customers
• Application scientists eg clinical research
  fellow
• Companies (eg pharma)
• Healthcare organisations

19
Scenario 3

• Cross validation
• Grid services providing image registration
• Reference datasets
• Definition of how to to add a new registration
  service or new data
• Can be used by researchers to compare algorithms
  prior to publication
• Challenges to compare algorithms on new problems,
  eg one brought by a pharma company, new
  algorithms

20
Scenario 0

• To build a grid application to enable comparison
  of registration algorithms
• Limited grid demonstrator
• Loose collaboration of resources, algorithms and
  researchers
• At least 3 algorithms
• At least 3 data repositories (?include
  pathology)
• At least 3 compute resources
• At least 3 research teams
• Demonstrate expandability, heterogeneity

21
Obstacles

• Image file formats
• We could use the DICOM core fields
• Access to grid infrastructure
• Hardware
• Support and training to set it up
• Architecture
• Coordination
• Portal
• registry

22
More Obstacles

• Transformation format
• Could initially provide tutorial, example
  standard transformation, i/o libraries
• Should we have a minimum subset of data eg
  vector field of resolution of original target
  image
• Origins etc
• How do you quantify differences between
  algorithms?
• Security and confidentiality
• We would start with anonymised data for which
  suitable consent had been obtained

23
Way forward

• We want to keep it small so that things can
  happen for scenario 0.
• We could keep it small by proposing it is
  initially done with no funding.
• This needs to be application lead
• The primary aim is to answer scientific questions
  about image registration

24
Partners and Roles

• INRIA linux cluster,
• already has a PhD student working on
  parallellising registration algorithm,
• could probably generate interest in putting this
  on the grid.
• KCL Imperial (IXI)
• Already implementing registration service
• NESC
• Could assist with registry and portal (a month or
  so of effort to do something)
• Hopkins
• Supercomputer on the grid running registration
• Others to give global coverage

25
Timetable

• Set up email list Sept 2003
• Invite key application groups to meeting at
  MICCAI 2003, Montreal in November 2003
• Initial demonstration of concept, and description
  of project for eHealth in January 2004
• Need a grid technology support from NESC and
  Clermont Ferrand