Use of Data Provenance and the Grid in Medical Image Analysis and Drug Discovery an IXI exemplar

1
Use of Data Provenance and the Grid in Medical
Image Analysis and Drug Discovery an IXI
exemplar

• Kelvin K. Leung1, Mark Holden1, Rolf A.
  Heckemann2, Nadeem Saeed3,
• Keith J. Brooks3, Jacky B. Buckton4, Kumar
  Changani3, David G. Reid3,
• Daniel Rueckert5, Joseph V. Hajnal2, Derek L.G.
  Hill1
• 1Division of Imaging Sciences, King's College
  London, UK
• 2Imaging Sciences Department, Imperial College
  (Hammersmith Hospital Campus), UK
• 3Imaging Centre, 4RA Disease Biology, ri-CEDD,
  GlaxoSmithKline, UK
• 5Department of Computing, Imperial College, UK

2
Overview

• Background
• Motivations
• Virtual data system
• Automatic delineation of multiple bones in serial
  MR images of joints in a disease model of
  Rheumatoid Arthritis (RA)
• Image registration and segmentation propagation
• Methods
• Prototype
• Results
• Conclusions

3
Motivations

• Medical imaging is going to play an important
  part in drug discovery
• Recent 76m investment by GlaxoSmithKline (GSK)
  and Imperial College on a new clinical imaging
  center
• Automatic analysis of medical image data
  requires
• Lots of storage space (each image is about 32Mb
  in this work)
• Computational power (running time is about 20-24
  hours for processing an image on a single desktop
  computer in this work)
• Motivated by the need of computational resources

4
Motivations

• The Grid has the potential to allow better
  collaboration between industry and university
  with the idea of virtual organisation
• University can provide image analysis algorithms
  as services to the industry, such as GSK, over
  the Grid
• Motivated by the need of better and more
  effective collaboration with the industry

5
Motivations

• Detail and reliable documentation of data
  provenance of all the analysis is very important
  in order to obtain regulatory approval for new
  drug.
• Part 11 of Guidance on industry issued by US Food
  and Drug Administration (FDA)
• Good Laboratory Practice (GLP) and Good Clinical
  Practice (GCP)
• Motivated by the need of data provenance

6
Overview

• Background
• Motivations
• Virtual data system
• Automatic delineation of multiple bones in serial
  MR images of joints in a disease model of
  Rheumatoid Arthritis
• Image registration and segmentation propagation
• Methods
• Prototype
• Results
• Conclusions

7
Virtual data system (VDS or Chimera)

• A system to enable documentation of data
  provenance, discovery of available methods and
  on-demand data generation (so-called virtual
  data)
• Developed by I. Foster, J. Vöckler, M. Wilde and
  Y. Zhao of University of Chicago
• It consists of
• A virtual data catalogue is a virtual data schema
  that provides a representation of computational
  procedures and their invocations.
• A virtual data language interpreter handles all
  the requests for constructing and querying the
  database entries.
• Data objects, such as input and output files, are
  described by logical file names (LFN), which are
  mapped to physical files via Globus replica
  catalog (RC) or Globus replica location service
  (RLS)

8
Virtual data system

• Virtual data language (VDL) is used to describe
  computational procedures and their invocations
• Computational procedures are defined by
  transformation (TR) statements. Example
• TR foo(input file1, output file2)
• Invocations are defined by derivation (DV)
  statements. Example
• To invoke foo with logical filenames file_a
  (input) and file_b (output)
• DV call_foo-gtfoo(file1_at_inputfile_a,file2_at_o
  utputfile_b)
• Virtual data schema allows the storage of TRs
  and DVs

9
Virtual data system

• Compound TR can be built so that workflow can be
  defined. Example
• To call foo twice and pass the output of the
  first call to the input of the second call
• TR compound_foo(input file_in, output file_out,
  io file_io)
• call foo(file1_at_inputfile_in,
  file2_at_outputfile_io)
• call foo(file1_at_inputfile_io,
  file2_at_outputfile_out)
• When requesting an output file from the system,
  an abstract DAG (contains only LFN) will be
  generated.
• A planner called Planning for Execution in Grid
  (Pegasus) converts the abstract DAG into a
  Condor DAGman script and submit it to the Globus
  universe of Condor.

10
Overview

• Background
• Motivations
• Virtual data system
• Automatic delineation of multiple bones in serial
  MR images of joints in a disease model of
  Rheumatoid Arthritis
• Image registration and segmentation propagation
• Methods
• Prototype
• Results
• Conclusions

11
Automatic delineation of multiple bones

• Rheumatoid Arthritis (RA)
• Is a chronic, systemic, autoimmune inflammatory
  disease.
• Targets synovial joints, in which there is a
  massive accumulation of blood-borne cells such as
  T cells and macrophages.
• Blood vessels are formed to support this new
  tissue and the whole mass is called pannus.
• Progressive erosion to cartilage and bone leads
  to disability in patients
• MR images were acquired in a disease model of RA
• Interested in the talus bone and the calcaneus
  bone in the ankle
• Delineate them from the MR images and study them,
  e.g. calculate volume to measure any erosion

12
Image registration

• Refers to the spatial alignment of two images so
  that corresponding features in the two images are
  matched
• The result is a spatial mapping or transformation
  that transforms positions from one image to
  positions in another image.
• Example Movie showing the rigid registration of
  two 3D MR images of a knee

13
Image registration

• Rigid registration translation rotation 6
  degrees of freedom (dof)
• Affine registration rigid skewing scaling
  12 dof
• Nonrigid registration warp one image into
  another one
• Very computationally demanding because of lots of
  dof
• Example Free form deformation (FFD) models local
  deformation as translation of a regularly spaced
  grid of points (control points)

14
Segmentation propagation

• Makes use of the spatial mapping calculated from
  the registration of two image to perform
  segmentation
• Requires an atlas
• An atlas is a reference image with labelled
  structures

15
Segmentation propagation
Atlas
calcaneus
All image analysis workflows were entered into VDS
Target image
Reference image
Manual segmentation of calcaneus
16
Overview

• Background
• Motivations
• Virtual data system
• Automatic delineation of multiple bones in serial
  MR images of joints in a disease model of
  Rheumatoid Arthritis
• Image registration and segmentation propagation
• Methods
• Prototype
• Results
• Conclusions

17
Prototype

• Simple web interface to replace some command line
  tools of VDS, Globus Toolkit 2.4 and Condor
• Researchers or clinicians working on medical
  image analysis may not be comfortable with
  command line tools and the virtual data language
• Developed using Java servlet on Apache Tomcat
• Web pages for
• Querying VDS for transformations and derivations
• Invoking transformations in VDS
• Querying, uploading and downloading files to and
  from Globus RLS
• Displaying job status in Condor

18
Prototype
Web portal machine running Apache Tomcat, Globus
client, personal Condor (job submission site)
Grid machine running Globus Gatekeeper, GridFTP
server, Globus RLS and Condor
Experimental condor pool of 4 machines
(storage and execution site)
19
Overview

• Background
• Motivations
• Virtual data system
• Automatic delineation of multiple bones in serial
  MR images of joints in a disease model of
  Rheumatoid Arthritis
• Image registration and segmentation propagation
• Methods
• Prototype
• Results
• Conclusions

20
Results
services
21
Results
Service to delineate the calcaneus and talus
from the target image
22
Results
23
Results
Jobs generated
24
Results
Job status in Condor
25
Results
Click to download files and view in vtkview
26
Results
Service to render the surfaces of the bones
27
Results
Job submitted
Job status
28
Results
29
Results
Browse all the executed services
30
Results
31
Overview

• Background
• Motivations
• Virtual data system
• Automatic delineation of multiple bones in serial
  MR images of joints in a disease model of
  Rheumatoid Arthritis
• Image registration and segmentation propagation
• Methods
• Prototype
• Results
• Conclusions

32
Conclusions

• We integrated Grid middleware and data provenance
  tool with medical image processing software in a
  prototype system with collaboration with GSK
• Data provenance of the results were kept in VDS.
  They can be queried and retrieved easily.
• Aim to satisfy guidelines issued by US FDA, GLP
  and GCP on the maintenance of audit trail of
  electronic records.
• The total processing time of delineating 12 bones
  from 6 subjects were cut down from about 132
  hours to about 33 hours (a factor of 4) by
  running the computing tasks on a Condor pool
  instead of on a single desktop computer

33
Further work

• More user feedback is required to evaluate and
  improve the system
• Further validation and application to a larger
  amount of subjects are required to determine the
  sensitivity of the delineation technique to
  disease progression

34
Acknowledgements

• EPSRC
• GlaxoSmithKline (GSK)
• Links
• IXI www.ixi.org.uk
• VDS www.griphyn.org/chimera