Speaker
Description
Summary
- The Medical Scenario
1.1. Pharmacokinetic Modelling
The pharmacokinetic modelling of the images obtained after a quick administration of
a bolus of extracellular gadolinium chelates contrast can have a deep impact on the
diagnosis and the evaluation of different pathogen entities.
Pharmacokinetic models are designed to forecast the evolution of an endogenous or
exogenous component on the tissues. To follow-up the evolution of the contrast agent
a sequence of MRI volumetric images is obtained at different times following the
injection of contrast. Each of these images comprises a series of image slices that
cover the body part explored. Since the whole process takes a few minutes, images are
obtained in different break-hold periods. This movement of the patient produces
artefacts that make images directly incomparable.
The study of pharmacokinetic models for the analysis of hepatic tumours is an
outstanding example of the above. A prerequisite for the computation of the
parameters that govern the model is the reduction of the deformation of the organs in
the obtained images. This process can be performed by co-registering all the
volumetric images with respect to the first one.
1.2. Co-registration
The co-registration of images consists on aligning the voxels of two or more images
in the same geometrical space by using the necessary transformations to make the
floating images as much as possible similar to the reference image.
In general terms, the registration process could be rigid or deformable. Rigid
registration only uses affine transformations (displacements, rotations, scaling) to
the floating images. Deformable registration enables the use of elastic deformations
on the floating images. Rigid registration introduces fewer artefacts, but it can
only be used when dealing with body parts in which the level of internal deformation
is lower (e.g. the head). Deformable registration could introduce unrealistic
artefacts, but is the only one that could compensate the deformation of elastic
organs (e.g. in the abdomen). Registration in 3D is necessary in this case, since the
deformation happens in the three axes. This is an extremely time-consuming process.
1.3 Post processing
Although the co-registration of images is a computationally complex process which
must be performed before the analysis of the images, it is not the only task that
needs high performance computing. Extracting the parameters that define the model and
computing the transfer rates for each voxel in the space will require large computing
resources.
The process of identification of the parameters consists on solving an
over-determined non-linear system of equations for each voxel on the image. This
process is computing intensive considering the amount of voxels of the images (on the
order of 6 millions per study). This process is highly parallel.
- Grid Application
In order to provide the necessary computing power to solve the co-registration and
the identification of the parameters, an application has been developed. This
application has been implemented considering the following points:
• Provide high performance. Be prepared to use a large Grid infrastructure to provide
the computational power.
• Usability. Reduce the complexity of the use of Grids by means of a user-friendly
interface. This interface must be open to its integration in other applications. The
choice is to implement a web-services based portal.
• Security. The program must deal with the risks of using remote resources.
Anonymisation is required and access control is very important.
• Reliability. The application must provide production capability, so assistance to
the Resource Broker must be provided on selecting the sites.
This Grid application uses a graphical user interface that calls web services that
implement the different tasks. This application is based on previous developments [2][3].
• Creating the proxy on the grid. Users have a certificate stored in the UI. The
private key is provided by the user and a proxy is created remotely on the UI.
• Transferring the data into the Grid. The user gets the images from local disks or a
scanner, transferring them to the FTP server of the UI. The data is copied to the SEs.
• Select the ranges of the parameters to test. The users can select the range of the
values of the input parameters (Maximum step length for the gradient descent
optimisator, Maximum number of iterations for the optimiser, Initial scaling factor
and Initial angle for deformation) that will be used for the jobs to be launched.
• Create the JDLs and define the arguments for the scripts of each job (One job per
registration and per combination of parameters).
• Run the jobs and retrieve the data to the Working Nodes.
• Monitoring of the evolution of a set of jobs.
• Downloading the output of all the jobs in a group with a single click. This implies
downloading the results, which were uploaded in the Storage Resources by the jobs, to
the UI, and from it back to the FTP server and the application.
-
Results
The results obtained can be considered in terms of performance and scientific
results. The results presented in this section are related to the images from a
clinical trial with 20 patients obtained at the Hospital Dr. Peset for this work.
Considering the performance, the required time to perform a registration of a
volumetric image in a PIII at 866 Mhz with 512 MB of RAM is approximately 1 hour and
27 minutes. Considering that the complete study performed involved 20 patients the
total cost would be 2331h 22m. Using a 20-procs computing farm the complete process
took 132h 50m. The computational cost using the Grid was 17h 35m.
The overhead of Grids is due to the use of secure protocols, remote and distributed
storage resources and the scheduling overhead, which is in the order of minutes due
to the monitoring policies which are implemented in a poll fashion.
Regarding the scientific results obtained. -
Conclusions and Future Plans
The computing requirements for a reduced clinical trial of 20 patients exceeds the
conventional computational capabilities of either a hospital or a research team.
Moreover, the need for computing is not constant and only after the clinical trials.
Thus, there is a need for a production platform with a high degree of reliability,
such as EGEE [1]. On the other side, the use of gLite is justified by the need for
access control in data and metadata and the improved metadata management.
The work is being completed with the implementation of the model parameters
computation on the Grid, which is also a time-consuming process and which could be
easily speed-up by the use of the Grid. - References
[1] EGEE. Enabling Grids in E-SciencE. http://www.eu-egee.org
[2] I. Blanquer, V. Hernandez, D. Segrelles, M. Robles, J. M. Garcia, J. V. Robledo,
“Clinical Decision Support Systems (CDSS) in Grid Environments”. From Grid to
HealthGrid (ISBN 1-58603-510-X), Vol: 112, pp: 80-89, 2005 IOS Press.
[3] Ignacio Blanquer, Vicente Hernández, Ferran Mas, Damià Segrelles, “A Framework
Based on Web Services and Grid Technologies for Medical Image Registration”. Lecture
Notes in Bioinformatics, ISSN 0302-9743, vol 3745, pp 22-33, Sptringer 2005.
