This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
The BOLD signal is small and contaminated by noise. Describe how data recording and analysis techniques overcome this problem to reveal significant task-related changes in brain activation
The size of BOLD signal changes, induced by local field potentials of neurons, are typically only a few percent. The sensitivity of detecting the change is dependent on the signal to noise ratio. Primarily, sources of contamination come from: system hardware, physiological processes, movement and unrelated cognitive processes. Noise impacts on both the temporal and spatial aspects of data acquisition and therefore techniques to predict and reduce data variability allow for more accurate detection and estimation of activation.
The MRI system predominantly contaminates the BOLD signal by: thermal noise, non-uniform magnetic field and scanner drift. Firstly, signal amplification by electronics can distort the amplitude of the signal, as a result of free electrons colliding with atoms converting energy to heat, effectively resistance. This is particularly detrimental to voxels that have a low signal to noise ratio. The resonant frequency of the atoms is dependent on field strength, therefore fluctuations impact upon the acquired signal, these inhomogenieties manipulate the MR signal both temporally and spatially. The field also has a tendency to gradually decrease known as scanner drift. The development of shimming coils and multichannel coil arrays has proved effective in homogenizing the field strength. Also, monitoring of field drift and spike noise have eliminated much of the instrumental instabilities (Weisskoff, 1996). Furthermore, effective field mapping using a dual echo time sequence acquires two images at each time point (Huxton, 2002) to compensate for variation over time. Arguably high field strengths have been used to increase the amplitude of the BOLD signal based on the theory that the MR signal increases quadrilaterally, whereas the field strength and thermal noise increase linearly (Turner et al. 1993). However the BOLD signal contributes only a fraction of the MR signal and so other sources of noise are also enhanced (Kruger and Glover, 2001). Hence improvement on detection and estimation of activation with a high field strength depends on successfully disambiguating physiological noise, which for complex higher processing tasks whereby lots of regions are activated may prove difficult.
Sources of physiological noise, for instance cardiac and respiration processes, produce signal changes that can resemble the BOLD signal associated with neural activity. However these changes are spatially variable (Birn at al., 2006), therefore temporal properties of fMRI can be used to extract the true BOLD signal. The sampling rate (TR) is very important in distinguishing signal from noise, because to obtain an accurate recording of the frequency of a phenomenon the sampling rate must be half that of the phenomenon. Where a TR of less than 1 second is used it is sufficient to sample most physiological noise and thus a temporal band filer can be applied, this will retain the signal of interest at the cost of physiological noise. However, T2* weighted images have a longer TR typically 2000ms and therefore this results in an under sampling, producing an uneven distribution of physiological ââ‚¬Ëœnoiseââ‚¬â„¢ over time. In this case, physiological noise contaminates the BOLD signal by coinciding with the spectrum frequency of the experimentally induced activations. Retrospective software such as RETROICOR (Glover et al., 2000) makes use of measuring physiological responses during the experiment and models these recordings onto the signal. However the equipment used to acquire physiological responses can be impractical and not appropriate for all experimental designs. An alternative approach makes use of sequencing the temporal data from the image (Frank et al. 2001). The sampling rate of an entire slice is relatively slow, but the rate at which data is actually being sampled is much faster, thus by converting each data point into temporal order the TR is sufficient enough to sample physiological noise. Although, this method only identifies noise temporally and even though physiological noises are considered to be relatively global in their effect (Lund et al., 2006), there are subtle spatial differences for example inferior regions of the brain are more susceptible to change than superior (Raj et al., 2001). Furthermore, Cheng (2010) made use of the finding that respiration has a linear relationship with fMRI phase signal and therefore assumes that respiration influences phase variations and so forth uses phase information to estimate the contribution of respiratory noise in the data. However, this approach still fails to appreciate spatial variations and the contribution of other aspects of physiological noise such as heartbeat. Additionally, the impact of physiological noise from other sources may co vary with respiration and similarly influence the phase signal. Therefore, psychological noise still poses a problem when aliasing neuron associated activations, this is largely due to the irregular spatial distribution of noise and the limitations of a long TR. Generally, cardiac and respiratory waveforms are very complex and co-vary with other physiological responses, for example heart rate varies with respiration, which varies to exertion. Further research could focus on the interactions between types of physiological noise and phase information.
Moreover, since each voxel has an absolute location, if the subject moves then subsequent changes in the voxels intensity can be mistaken for the BOLD signal. Co-registration aims to normalise volumes so each voxel is in the same spatial location throughout all volumes. However, there are more than 15 billion possible combinations of movement parameters at each time point, therefore the most common correction technique assumes a rigid head transformation and predicts the best fit of voxels to a reference image. The process estimates a correction and then calculates the cost function (intensity difference between the reference and correction) for that correction and small deviations from it, then selects the translation with the smallest cost function. This local minima testing may fail to return the best solution if it is further afield and so one should manually fit the data before applying computerised analysis. Also, co-registration is conducted by mutual information and therefore it is wrong to interpret corrected images as voxels that match location across volumes, but that the reference image can predict to an acceptable level the intensity of voxels in each volume. This has particular importance when identifying regions of activity.
A solution to this problem is proposed by Thacker et al. (1999). They validate the motion correction of computer software, by plotting a scatter graph of the local spatial correlation scores of active voxels against the local field gradient in a region. This allows a visual representation of how broad the distribution of values in a region of interest is. If co-registration has been accurately performed then correlation scores should cluster along the local image gradient axis. This is a useful technique to check the software motion correction, but there still remains the question of how to improve motion correction methods. An alternative approach may be to make use of external monitoring, where an infra-red tracking device measures head motion directly (Tremblay et al., 2005). The coordinates of the head can be calculated at any time point and used retrospectively to adjust images for movement accordingly and also used in real time to inform researchers of the impact of movement on the current data being acquired. Nevertheless, a source of motion comes from the tissues within the brain, for example from pulsation in ventricular fluid spaces (Dagli et al., 1999), this is often ignored in analyses, but further research could explore higher order algorithms to take this into consideration. In addition, error associated with motion correction can be included into the general linear model as a confound, but often motion is associated with stimuli presentation and so this risks removing task-related activity.
Secondly, movement impacts on spatial analysis by creating image geometric distortions that make it difficult to achieve an accurate registration, corrections using the linear models are limited to scaling, translation and rotation, however distorted images introduce non linear terms in the direction of the distortion. Additional problems with localisation come from contamination of large surface vessel signals. The BOLD signal changes in these can be between 10-20% (Ogawa & Sung, 2007), especially when closely activated areas transfer their oxygenated venous blood to a draining surface vessel, these large values distort the overall BOLD signal. Furthermore, the small parenchymal venules that are in close proximity to active sites are overpowered by the larger draining veins that are distant from the active parenchyma. This is commonly known as the brain or vein argument (Fahm et al 1994). Krings et al. (1999) made use of the time-series of task induced activation to disambiguate activation of smaller paranchymal vessels and larger draining vessels. Nonetheless, this technique may prove useful when identifying a single region of activity in relation to task-induced activation, for example for pre-surgical planning, however this technique fails when investigating activation networks or resting state fMRI. Further research could explore how to mask these large vessels using Fluid Attenuated Inversion Recovery (FLAIR) (Bailey, 2007).
Additionally, researchers aim to disambiguate task-related activations from unrelated neural processing. One basic strategy to reduce noise in the signal is to average activations over many trials. The principle being activations associated with the task are systematic and noise is random, therefore averaging will enhance the power of scores that vary together and reduce those which are inconsistent. A single subject participating in many trials raises issues with habituation, fatigue and loss of attention, therefore averaging across subjects can be beneficial. However additional consideration should be given to variation in the BOLD signal and unrelated neural activity between subjects, for instance factors such as age (Garrett et al, 2010) and between patients and normals. Also, differences in cognitive strategy, where sometimes the strategy change is distinct enough activation may represent a combination of the different regions used in each strategy within the same trial. Averaging trials in blocked designs fail to differentiate activation changes during the block, therefore if changes in strategy or brain activation during the trial are anticipated event related designs may provide a less contaminated BOLD signal. Due to their time locked properties event related designs are adept at showing transient activations, nevertheless for other hypotheses that manipulate individual stimuli in relation to a context then block designs are better suited to establish sustained activation over a trial. Furthermore, event related designs have the benefit of acquiring two data sets that are measured at odd and even measurements, therefore obtaining 1s TR recordings for a 2s TR scanner, which aids better aliasing of BOLD signal as discussed earlier. Therefore to reduce unrelated BOLD signal, experiments need to be designed to suit the activation changes that relate to specific hypotheses under investigation.
Similarly, spatial smoothing techniques make use of averaging to reduce noise. Using the assumption that the BOLD signal from neuronal activity will not vary greatly between nearby voxels, but noise will, voxels are averaged over a small area. Smoothing is beneficial when comparing across subjects as it is unlikely they will have exactly the same voxel location of activation, but it is debatable what the appropriate filter size is, as an overly large filter may render localisation inaccurate. Caution is also needed because if noise in local voxels is correlated, for example in areas susceptible to noise such as near air spaces or CSF cavities (Nencka and Rowe, 2007), then averaging will increase noise in these voxels.
In conclusion, sources of noise that still produce controversy are predominantly physiological and motion. This is primarily due to their variations over both time and space. Further research could focus on how sources of physiological noise co-vary and their relationship with phase information (Frank et al,) algorithms to account for movement of brain tissues and solving the brain-vein controversy. More importantly when concluding the findings of an experiment, the potential impact noise could be playing should be considered and evaluated when relating areas of activation to function and previous research.