You are here

Home

Historical Interest Only

This is a static HTML version of an old Drupal site. The site is no longer maintained and could be deleted at any point. It is only here for historical interest.

Gaussian Process deconvolution for perfusion imaging: evaluation of the usage of distributed and parallel computing

17 November 2008 - 9:31am — Jano.van.Hemert
Student: 
Fan Zhu

Final version of the thesis submitted.

Original project description:

Deconvolution is used in perfusion imaging to obtain the impulse residue function (IRF) that is then used to create parametric maps of relevant haemodynamic quantities such as CBF, CBV and MTT [1]. A popular method to achieve this is Singular Value Decomposition (SVD), but it has been shown that for MRI Gaussian Process Deconvolution (GPD) [2] is comparable to SVD when determining the maximum of the IRF, and superior estimating the full IRF. Furthermore, it clearly outperforms SVD when the signal-to-noise ratio improves. Gaussian Process regression [3] arises from a Bayesian approach to the regression problem, and as in the case of other kernel-based methods the scalability with data size is very poor. This constitutes the main drawback of this technique to compute deconvolution when compared with SVD.

The currently running Wyeth-TMRC multicenter project on acute stroke brings the opportunity to test this technique with data from several SINAPSE centres and different modalities. This PhD project will benefit from the expertise in these centres and would seek to collaborate with them through the centres’ contacts: M.J. McLeod (Aberdeen), J. Wardlaw (Edinburgh) and K. Muir (Glasgow).
The project will research the possibilities that distributed (and parallel) computing brings to make this method usable in practice. There have been some previous works like the dataparallel approach proposed in [4]. The project will study the consequences on the final results of the local learning used in it. As a by product, the project will produce a data processing framework prototype reusable for other types of image processing.

Project status: 
Finished
Degree level: 
PhD
Background: 
MSc in Computer Science essential. Strong background in imaging and distributed computing important.
Supervisors @ NeSC: 
Jano.van.Hemert
David.Rodriguez
Malcolm.Atkinson
Other supervisors: 
Prof Joanna Wardlaw (SF Brain Imaging Research Centre, University of Edinburgh) Dr Trevor Carpenter (BRIC, University of Edinburgh)
Subject areas: 
4
Projects: 
SINAPSE
Student project type: 
PhD student project
References: 
1. Ostergaard L, Weisskoff RM, Chesler DA, Gyldensted C and Rosen BR “High resolution measurement of cerebral blood flow using intravascular tracer bolus passages: I.Mathematical approach and statistical analysis” Magn. Reson. Med. 36 715–25 2. Andersen IK et al; “Perfusion Quantification Using Gaussian Process Deconvolution”. Magnetic Resonance in Medicine 48:351-361 (2002). 3. Williams CKI and Rasmussen CE; “Gaussian processes for regression”. Advances in neural information processing systems, (1996), 514-520. 4. Choudhury A, Nair PB and Keane A; “A Data Parallel Approach for Large-Scale Gaussian Process Modeling”. Proc. the Second SIAM International Conference on Data Mining (2002).

      Creative Commons License