The detection, mapping, and quantification of neural activation with FMRI relies on extracting information from noisy measurements. In any such problem, it is necessary to understand both the signal that is to be detected, and the noise that interferes with the detection process. I will discuss some characteristics of the FMRI signal and noise, and will explain why both parts of the problem are active areas of research.

When one knows the form of the signal, then detection of activation is essentially a pattern matching problem: which parts of the brain show the expected pattern, and which do not? The most widely used method for this is a least squares fit technique. When there is no expected form to the received signal, then detection of activation involves looking for unknown temporal patterns. There are two methods widely used for this: principal components analysis, and fuzzy clustering.

Characterizing the response of the MR signal to changes in brain tissue state is one of the most active areas of research. Understanding, detecting, and correcting for the many artifacts that can occur in MRI time series of the human brain is also important. These include scanner effects, subject motion, and spatial correlation of the ``noise" from physiological fluctuations in the subject.

If each voxel is considered separately, so that any possible spatial activation pattern is allowed, then a large (> 10000) number of activation decisions must be made -- this is the ``multiple comparison" problem. To ensure that few of these decisions produce false activations, the threshold for detection must be made very stringent, making the detection of weak activations unlikely. One way to overcome this is to restrict the allowable spatial activation patterns; for example, to detect only in clusters above a certain volume.

Back to Workshop Schedule