Common types of artefacts
Spikes. These occur due to electrical instability e.g., static electrical discharge. They appear as regular patterns of stripes across an image.
Ghosts. These occur due to phase offset in K-space, or periodic motion due to breathing or heartbeat. They appear as ghost images to the side of the brain. They can cause activations to show up outside of the brain, or mislocation of activations within the brain.
Artefacts are best spotted by human eye, using time series animations. Indepedent components analysis is also useful for identifying artifacts such as those caused by within scan head motion etc.
Artefacts caused by EPI
Dropout. There is an inhomogenity in the main magnetic field (B0) at air-tissue interfaces, which often causes loss of signal. This occurs at regions close to sinuses such as the orbitofrontal cortex and lateral temporal lobe. This can give misleading results that these regions are not active in tasks when they actually are. The best way to spot this is by overlaying images with structural ones.
Geometric distortion. When gradients are applied to encode spatial information, there is often distortion due to inhomogenities. This often occurs in the anterior pre-frontal cortex and the orbitofrontal cortex.
These are used to correct artefacts caused by magnetic field inhomogenities
For most data, images are collected one slice at a time - either in ascending/descending order.
Alternatively they can be collected through interleaved acquisition where every other slice is acquired sequentially.
Image shown below is for a repetition time of 2seconds.
Slice time correction
Due to 2D image acquisition where images are collected one slice at a time, there are systematic time differences for acquisiton of different parts of the image. Time differences depend on repetition time. However, the model that is applied to the data assumes they were all collected at the same time, creating a mismatch.
Slice time correction interpolates all the data to match the timing of a reference slide.
Practice is shifting away from this though, as it can result in artefacts from one image being propogated through an entire time series. Instead, TR
Stimulus correlated motion
This is motion that is correlated with task states. For example, tasks requiring overt speech or large muscle group movements may show task motion. Movement could also increase in difficult tasks due to frustration. Cognitively impaired subjects and children commonly exhibit these task related motions.
Because the activity is correlated with the task, it can result in artifactual task related activations. Furthermore, removal of these signals may also remove task-related signals.
Correlation between movement and the task can be reduced using jittered-event designs, due to the delayed nature of the bold response.
Benefits of smoothing
Increasing signal-to-noise ratio. The gain in signal for larger features may outweigh the cost of loss in signal for smaller features. It can also reduce dropout, and overcome increased noise due to small signal size.
Overcoming interindividual varibility. There are differences in the spatial location of functional regions that cannot be corrected by spatial normalization.
Some statistical methods require smoothing. E.g., Guassian random fields.
This is the most common method of smoothing, where the 3D image is convolved with a 3D kernal. The amount of smoothing is determined by the width of the kernal (stated in terms of full width half maximum - FWHM). The larger the FWHM the greater the smoothing.
Kernals should be no larger than the size of the activations. Twice the voxel dimensions is standard.