|
Magnetic resonance imaging (MRI) is a powerful technology forobtaining high quality images of the human brain (and otherorgans). MRI generally involves no damaging chemicals or radiation,unlike other technologies such as x-ray, CT or PET. The patient liesinside the scanner for 5-30 minutes, and a finely detailed 3D structural image is produced, showing different brain structures,including diseased or damaged areas.Functional MRI (FMRI) uses MRI in a slightly different way. Insteadof a single structural image, the scanner is used to take a seriesof much coarser 3D images - for example, obtaining an image every 3seconds for 10 minutes. The scanner is tuned so that these images aresensitive to brain activity. It is therefore possible to compareimages taken when the brain is active (for example, when viewing aflashing light), with images taken when the brain is resting. Thebrain areas which change in brightness correspond to the areas usedfor the particular activity (for example, showing the vision brainarea). Thus we can study, in fine detail, which parts of the brain areused for which purposes, to learn about the functioning anddevelopment of the healthy brain, and also about brain disease anddamage.However, there are serious limitations to the questions that cancurrently be asked using FMRI. FMRI data is noisy ; the signal in thedata due to real brain activation is often quite small, compared withthe part due to noise from the MRI machine, and factors such asheartbeat or breathing. FMRI analysis research and software attemptsto separate out the real activation signal from the noise. Unless thisis done well, only the simplest, least interesting FMRI experimentscan be carried out. However, both the noise and the activation signalare very complex; we therefore need sophisticated methods tounderstand the data fully, and to see what is going on inside thebrain.Previously, we developed advanced mathematical predictions of how thesignal and noise change over time. This model-based approach is easyto use, but is unable to accurately find the most complexunpredictable noise, particularly that due to effects such asheartbeat or head movement. In other work, we developed asophisticated approach to finding such complex noise and activationsignal. This model-free approach looks for consistency in anartefact or signal across both time and space. It has been shown to bepowerful at finding complex noise, but does not always find the brainactivation, if the signal is weak, or limited to a small brain area,or if the brain is doing several things at the same time.These two approaches are clearly complementary; however, no-one hasyet put them together into a single, sophisticated, all-in-onemethod. In this project we propose to do this - to take the bestaspects of space+time signal+noise modelling, and fuse this withautomatic, model-free, identification of the complex noiseartefacts. We will enhance the model-free methodology with some otherprevious work on categorisation of brain areas into non-activating andactivating ( spatial mixture modelling ) and then integrate thisinside a complete model-based framework, where each part informs theother of different brain and noise components in the data.In order to achieve all of this in a way that is as foolproof andsensitive to brain activation as possible, we will need to useadvanced mathematical techniques, to get the best possible results. Weaim to develop FMRI analysis software which we can distribute to MRIcentres around the world, enabling medics and researchers to learnmuch more about the brain than is currently possible.
|