Functional Magnetic Resonance Imaging (fMRI) is a neuroimaging technique used to study brain functionality to enhance our understanding of the brain. This technique is based on MRI, a painless, noninvasive image acquisition method without harmful radiation. Small local blood oxygenation changes which are reflected as small intensity changes in the MR images are utilized to locate the active brain areas. Radio frequency pulses and a strong static magnetic field are used to measure the correlation between the physical changes in the brain and the mental functioning during the performance of cognitive tasks.
This project presents approaches for the analysis of fMRI data. The constrained Canonical Correlation Analysis (CCA) which is able to exploit the spatio-temporal nature of an active area is presented and tested on real human fMRI data. The actual distribution of active brain vessels is not known in the case of real human data. To evaluate the performance of the diagnostic algorithms applied to real human data, a modified Receiver Operating Characteristics (modified ROC) which deals with this lack of knowledge is presented. The tests on real human data reveal the better detection efficiency with the constrained CCA algorithm.
A second aim of this project was to implement the promising technique of constrained CCA into the software environment SPM. To implement the constrained CCA algorithms into the fMRI part of SPM2, a toolbox containing Matlab functions has been programmed for the further use by neurological scientists. The new SPM functionalities to exploit the spatial extent of the active regions with CCA are presented and tested.
Interaction Between Engineering and Medicine
Cross-disciplinary activities in the field of medicine are of a growing interest. Biomedical engineering is a discipline that joins the knowledge in engineering sciences with clinical practice. New devices, algorithms, processes and systems are developed to improve the medical practice and health care. A researcher in the field of biomedical engineering not only needs to be familiar with the relevant applications of engineering in medicine but also with the basic life sciences.
Medical Signal Processing and Neurology
Medical care incorporates diagnostic, monitoring and therapeutic issues. Typically relevant patients characteristics are measured, interpreted and an appropriate decision about the therapeutic action is taken. The medical signal processing focuses on the extraction of the useful information from medical images and signals.
Hydrogen is the most abundant element in the human body (especially in water and fat). We consider a proton of a hydrogen atom to understand how particles with spin behave in a magnetic field. The magnetic moment of this proton causes the proton to behave like a tiny magnet with north and south poles.
Placed in an external magnetic field B0 of a magnetic resonance scanner, the spin vector aligns itself with the external field like a magnet would. There are two possible energy configurations, a low and a high energy state where the spin vector is aligned parallel or anti-parallel with respect to the external magnetic field B0 as demonstrated by figure 3.2). Slightly more protons are in the low energy configuration, giving a net magnetic field in the same direction as the external magnetic field.
We cannot observe directly changes in brain activation due to a stimulus. But already in the end of the 19th century were local blood flow changes in active brain areas predicted. All neurons in the brain consume oxygen. The haemoglobin molecules in the blood provide the neurons continuously with new oxygen. An increase of neuronal activity increases also the demand of oxygen. This leads to an increased concentration of oxygenated blood in the capillaries surrounding the active brain area.
Spatial filters are used to take the spatial extent of the active region into account. This filters create weighted averages over the voxel values in the neighbourhood of the analysed voxel time series.
- Gaussian filters
- Steerable filters
The most trivial BOLD response model is a boxcar model. The boxcar model has only one temporal basis function. The parameters we need to know are the sampling frequency (in Hz ) the number of points per interval of resting state and activity period, respectively as well as the number of such intervals of resting and work.
Figure 5.4: a) shows different BOLD model shapes. Realistic shapes lie around the main basis function y1( t ). The change of variable trick is applied. Now the weights can be constrained to non-negative values giving realistic shapes lying in the first quadrant in b)
In a general case the active regions have arbitrary orientations and can thus be captured better if we can adapt our model shape to different orientations and not just to the spatial extent as in the scale adaptive model. For this reason we use steerable filters (see figure 5.7).
All of the presented new functions enable the user to apply the constrained CCA technique to the fMRI data using SPM. In this section the new functionalities of SPM are presented. At any voxel of the MIP plot the corresponding time-series and the fitted response model function can be displayed (see figure 6.4). The active region can be overlaid to a selected volume scan to visualize the active region (see figures 6.5).
It is always preferable to work with simple models and methods with good performances and computational efficiency. In the field of analysis of functional MRI data, the widely used GLM method has such properties. This project shows that the constrained CCA method is an analysing technique with even better properties. Spatial geometry of the active brain area is respected, unrealistic model shapes are rejected and the computational time stays within acceptable limits.
Even if there exist a large number of neural activity detection methods, GLM is the most commonly used one in practice. The software tool SPM is also GLM based. CCA is in fact a natural extension of the GLM technique. GLM can be considered as an unconstrained CCA with just one spatial basis filter, usually a Gaussian smoothing filter. The implementation of constrained CCA was the second part of this thesis.
The neural activity detection method based on constrained Canonical Correlation Analysis (CCA) has been presented and was tested on real human fMRI data. It has been shown that this method is an extension of the General Linear Model (GLM) analysis technique widely used to find neural activity in fMRI time series. It has been shown how realistic temporal basis functions as well as spatial filter basis functions can be constructed to improve the efficiency of the CCA technique. For testing the efficiency of fMRI detection methods applied on real human data, a modified Receiver Operating Characteristic (ROC) method has been presented. In the results chapter it has been shown how all these concepts improve the neural brain activity detection performance.
The implementation of these novel parts into the Statistical Parameter Mapping (SPM) software has been presented. The new features introduced for the constrained CCA have been illustrated and tested.
Source: Linköping University
Author: Sabina Breitenmoser