Basser Seminar Series : Postgradute Research Seminar Series

Unsupervised classification of dynamic positron emission tomography datasets

Brian Parker
Postgraduate Student
Biomedical and Multimedia Research (BMIT) Group
Multimedia Computing Research Laboratory

Wednesday 31 August 2005 2-3pm

Basser Conference Room (G92) Madsen Building

Abstract

Positron emission tomography (PET) is a nuclear medicine modality that images injected radioactive tracer molecules to provide information about biochemical processes in the body. Dynamic PET imaging involves multiple snapshots taken over time, allowing for analysis of the time-varying behaviour of the tracer.

In this talk, I will present some results from my PhD on the problem of functionally classifying and analysing noisy dynamic PET datasets to extract weak signals, and in particular the extraction of cranial arterial and venous blood vessels.

The dynamic brain PET datasets are first preprocessed using multivariate analysis techniques, and then a segmentation/clustering using a Mumford-Shah energy is used to classify the dataset into functionally distinct regions. The Mumford-Shah model includes a region-compactness term, and the hypothesis that that this term improves demarcation in noisy datasets is experimentally validated by comparison with clustering approaches without such a term.