How can we model the brain and its PET images
Lead Research Organisation:
University of Warwick
Department Name: Statistics
Abstract
The aim of this project is to explore and develop a novel statistical methodology for analysis of neuroimaging data. Specifically, we will look to implement Sequential Monte Carlo approaches towards understanding Positron Emission Tomography (PET) images of the brain. Though this imaging modality produces large data sets, current analysis processes impose inaccurate assumptions or utilise approaches that may result in loss of spatial resolution which may not produce the most effective models. Thus a primary objective would be to incorporate spatial information and produce accurate and reliable models of the PET brain images. Ultimately, the methodology would need to balance the challenges of generating realistic models and being computationally tractable. In the fullness of time, a further secondary objective would be to extend analysis to multiple subjects and produce hierarchical models.
Organisations
People |
ORCID iD |
| Adam Johansen (Primary Supervisor) | |
| Denishrouf Thesingarajah (Student) |
Publications
Thesingarajah D
(2022)
The node-wise Pseudo-marginal method: model selection with spatial dependence on latent graphs
in Statistics and Computing
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509796/1 | 01/10/2016 | 30/09/2021 | |||
| 1943006 | Studentship | EP/N509796/1 | 02/10/2017 | 31/03/2021 | Denishrouf Thesingarajah |
| Description | Using the developed novel statistical algorithm it is possible to produce more accurate images of the volume of distribution of tracers from Positron Emission Tomography (PET) images of the brain. This has been accomplished by including spatial dependencies between the sub-units, or voxels, of the images within the statistical analysis of the PET image. Typically, in most current statistical techniques for analysis of brain PET images, there is a assumption that each voxel is independent. This assumption is made purely for computational reasons, and the dependence due to the complex structures seen in brain are not reflected in the statistical analysis. By using a Bayesian approach, together with restricting the dependence to the model order level, the computational algorithm developed as a result of this work allows for the incorporation of spatial dependence. Further more, at the current stage of the research (the award is still active), using a approximation of the proposed method add very little extra computational costs when compared to assuming spatial independence. The non-approximation version of the method has a high computational costs, so we will now focus on reducing this costs and/or showing that the approximation gives relatively reasonable results. Other smaller outcomes achieved thus far include: i) implementation (software) of the proposed algorithm, together with two other Bayesian methods that could be extended and published for use by other researchers. ii) some empirical studies and experiments (and further studies to be done in the near future) of the above algorithms for evaluating performance and robustness. Including more in depth designs of experiment --- which are lacking in current studies of other methods of PET image analysis. Currently we are the final stages of evaluating the performance of this algorithm on real observed data sets. |
| Exploitation Route | For researchers, and possibly clinicians, that use PET images, the proposed statistical method could provide a more accurate image of the volume of distribution. It also allows for the incorporation of spatial dependence and for studies of the structural aspects of PET images. More generally, the proposed method could also be used by statisticians or data scientists to perform model selection in any context where there are latent (discrete) graphs. For example this could include contexts where compartmental models are used and there is a requirement to include spatial dependence. The method can also be further extended: i) To be more computationally efficient, for example by using adaptive methods of tuning parameters; ii) To incorporate prior knowledge or data of the brain structure, for example using other structural imaging modalities such as computed tomography (CT). iii) To be possibly be used in conjunction with existing or novel region of interest (ROI) techniques. |
| Sectors | Healthcare,Pharmaceuticals and Medical Biotechnology,Other |
| Title | bayespetr |
| Description | A R(Cpp) package that allows for analysis of dynamic PET image using Bayesian statistics. |
| Type Of Technology | Webtool/Application |
| Year Produced | 2021 |
| Open Source License? | Yes |
| Impact | - Allows research to apply Bayesian statistical analysis to dynamic PET images. |
| Description | Youn Researchers Meeting Talk. |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Local |
| Primary Audience | Postgraduate students |
| Results and Impact | Presentation followed by discussion of research with other PhD students. |
| Year(s) Of Engagement Activity | 2019 |