Derek Hill

1

• Derek Hill
• KCL, Imperial, Oxford
• http//www.ixi.org.uk

2
Team

• Derek Hill, Kelvin Leung, Bea Sneller, Jinsong
  Ren, Julia Schnabel, Jason Harris KCL
• Jo Hajnal, Daniel Rueckert, Michael Burns, Andrew
  Rowland, Rolf Heckerman, Carlos Thomaz, Imperial
• Steve Smith, John Vickers, Oxford

3
Information eXtraction from Images (IXI)

• 3 year UK e-science project funded by core
  programme
• Additional support from GSK, Philips Medical
  Systems, Dunhill Charitable Trust
• Uses grid-enabled image registration and
  segmentation for drug discovery, medical
  research, and decision support in healthcare.

4
Image registration
Reference image (example slice)
Database subject image (example slice)
5
Brain image segmentation
6
Application to large cohorts
Example slices From MRI Volume images
7
Tissue probability maps
Tissue probability maps generated on 16 CPU
parallel computer at Oxford
Grey matter (coronal plane)
8
The IXI consortium

• Kings College London (Derek Hill, Dave Hawkes,
  Steve Williams, Gareth Barker)
• Imperial College (Jo Hajnal, Daniel Rueckert,
  David Edwards, LESC)
• University of Oxford (Steve Smith)
• Jointly coordinated by Derek Hill Jo Hajnal
• 9 full time researchers funded by the project

9
Research activities

• Image acquisition and analysis
• Between all sites have about 100 full time image
  analysis researchers (students and post-docs)
• We distribute various image analysis s/w,
  including image-registration.com (KCL) and FSL
  (from Oxford)

10
(No Transcript)
11
Why IXI?

• We call this project Information eXtraction from
  Images to emphasize the key concept which is
  using image analysis to generate image metadata
  information about the images and the generic
  applicability of this technology.

12
Why the grid?

• Data grid
• Sharing distributed image databases
• Enables collaborative working
• Compute grid
• on demand computing provided by distributed
  infrastructure
• Users can access high performance computing when
  they need it
• Algorithms presented as grid services that can be
  combined with workflow tools
• Provenance tools (eg Chimera) to provide
  electronic paper trail evolving link with
  Wilde/Foster Argonne National Lab
• People in virtual organizations
• Researchers can work together more effectively
• New ways for industry and academia to collaborate

13
Technical aims

• Scalability
• To show that the grid can scale medical image
  analysis to huge cohorts, using condor between
  sites
• Ability to share data across sites
• Interoperable databases
• Secure file transfer to trusted machines
• Grid services for image analysis
• Wrap image analysis algorithms to create grid
  service
• Provenance
• Keep track of how all results were obtained
• Information Extraction methodology
• New algorithm that take advantage of the grid

14
Exemplars

• Developmental neuroimaging
• Neonates from Hammersmith
• Children/teens from Institute of Psychiatry
• Drug discovery
• Pre-clinical brain and joint imaging
• Decision support in healthcare
• Normative reference data in dynamic brain atlas
• Cardiac MRI dynamic image analysis

15
Normative MRI reference data

• 600 normal subjects, approximately uniformly
  distributed between 18 and 80
• T1 volumes, multislice spin echo, angio and DTI
  on sub-cohort
• medical history questionnaire
• 1.5T and 3T scanners, different vendors
• Ethics approval for sharing on grid

16
Example use of normative data dynamic atlas
construction

• For any patient, identify nearest 20 reference
  subjects by age and epidemiological similarity
• Construct customized atlas from reference
  subjects to assist interpretation

17
Workflow of busy radiologist
Load patient image from worklist
Easy?
Yes
No
diagnosis
18
need reference data
19
200 reference subjects
Example slices From MRI Volume images
20
KINGS COLLEGE LONDON
IMPERIAL COLLEGE
Kings College London (Guys Campus)
Oxford University
21
KINGS COLLEGE LONDON
IMPERIAL COLLEGE
Kings College London (Guys Campus)
Oxford University
22
KINGS COLLEGE LONDON
IMPERIAL COLLEGE
Kings College London (Guys Campus)
Create atlas
Oxford University
23
The Radiologists view
24
(No Transcript)
25
(No Transcript)
26
(No Transcript)
27
Conclusions

• The dynamic atlas provides a customized
  authoritative reference presented in an intuitive
  way
• The doctor can see at a glance the normal range
  of sizes and shapes of each brain structure,
  overlaid on the patients own scan, assisting
  diagnosis.
• The grid will bring new ways of working to the NHS

28
Achievements

• Wrapping of image registration algorithms from
  within our consortium and also from a group at
  INRIA in France for demonstration of grid-enabled
  cross-validation of algorithms (demonstration at
  HealthGrid 2004,Clermont- Ferrand)
• Testbed based on XML workflow schema providing
  web access to grid services
• Use of IXI components to delineate talus and
  calcaneus from wrist to quantify disease
  progression in model of rheumatoid arthritis
  (collaboration with GSK) Paper presented at
  IEEE ISBI conference, April, USA

29
Architecture for intraoperatible image
registration (health grid demo)
Web-based portal
Local client
INRIA MPI Cluster
Images on local client
Imperial Condor Cluster
Globus
30
(No Transcript)
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
IXI testbed

• Resources
• 400 node sun grid engine cluster, London
  e-science centre
• 200 node condor installation, Imperial College
• 45 node condor installation, KCL
• Distributed image database, 3 sites (MySQL based,
  directly connected to MR scanners for data
  acquisition at 2 sites)
• globus installed at each site

35
IXI test bed system design

• xml schema language to describe existing image
  analysis applications
• Defines common types, parameters, i/o of each
  component, relationships between input and output
• Defines categorisation information for
  application discovery
• Used to construct image analysis workflows

36
IXI testbed Workflow Service

• OGSI compliant GT3 service, executes workflow
  based on xml schema
• Maps workflow to RSL specification or grid
  service invocation
• Handles dependencies between each workflow stage
• Tries to execute as much of workflow in parallel
  as possible.

37
IXI testbed service discovery

• OGSI based registry deployed at each site
• Users can register applications that they wish to
  make available to the project
• Registries aggregated to project-wide registry,
  which can be queried by user

38
IXI testbed Example Application

• demonstrator
• Database can be queried for head scans (one
  selected as reference) which are accessed by the
  workflow engine using grid-ftp
• Each head passed through workflow to extract
  brain
• All images aligned with reference
• Atlas of variability produced
• Accessible via a web server for users without
  globus installed
• Aim to demonstrate easy of analysis for
  non-expert users.

39
Drug discovery with provenance

• Pharmaceutical industry in investing massively in
  imaging (eg 70m investment at Imperial
  announced last month)
• For drug discovery, keeping track of exactly how
  result were obtained is critical
• We use the Virtual Data Systems Chimera system
  within a web interface to do this

40
Application - drug discovery

• Disease model of Rheumatoid Arthritis (RA)
• Injected with disease inducing agent
• MR images were acquired
• Interested in talus and calcaneus
• Identify them from the MR images and study them,
  e.g. calculate volume to measure any erosion

41
Segmentation Propagation
Rigid non-rigid registration
calcaneus
Target image
Reference (atlas) image
Displacement field
Apply displacement field
Computed boundary of calcaneus
Manual segmentation
42
IXI provenance system

• Web interface wrapped around VDS, Globus Toolkit
  2.4 and Condor
• Tomcat (https), VDS, Globus client, Condor on my
  machine
• Web portal
• Globus gatekeeper, GridFTP server, Globus RLS,
  Condor on another machine
• Storage site and execution site
• Not yet integrated with IXI testbed

43
My system
services
44
My system
Service to delineate the calcaneus and talus
from the target image
45
My system
46
My system
Jobs generated
47
My system
Job status in Condor
48
My system
Click to download files and view in vtkview
49
Result intra-subject registration
Day 3
Overlay images with the computed boundaries of
calcaneus highlighted
50
Result inter-subject registration
Day -12
Overlay images with the computed boundaries of
calcaneus highlighted
51
My system
Service to render the surfaces of the bones
52
My system
Job submitted
Job status
53
My system
54
My system
Browse all the executed services
55
My system
56
Provenance requirements

• Access control and security
• We have some unusual provenance requirements
• Provenance information needs access control so
  not everyone can see provenance of data
• We have started a collaboration with Mike Wilde
  and Ian Foster using our application as a use
  case for VDS.

57
Points for Discussion

• Web interfaces lack features
• Need more sophisticated queries
• Derivations are very specific
• Better searches needed (eg with wild cards)
• Better user interaction
• Repeating analysis eg with new version of s/w or
  new reference data
• Better ways of defining workflows (eg drag and
  drop workflow components)

58
Conclusions

• Medical image analysis has some characteristics
  that make it well suited to grid computing
• Algorithms have increasing computational
  complexity (gt moores law)
• There is a need to deal with larger data volumes
• Latency is not critical
• Collaboration is essential
• Regulatory environment requires good curation and
  provenance