Comparison of Robarts s 3T and 7T MRI Machines for obtaining fmri Sequences Medical Biophysics 3970: General Laboratory

Transcription

1 Comparison of Robarts s 3T and 7T MRI Machines for obtaining fmri Sequences Medical Biophysics 3970: General Laboratory Jacob Matthews 4/13/2012 Supervisor: Rhodri Cusack, PhD Assistance: Annika Linke, Postdoctoral Fellow

2 Introduction MRI has been a rapidly expanding and evolving imaging modality in recent years. This is because it is one of few non-invasive, non-ionizing forms of gathering information about the structure and behavior of our internal organ systems. This is doubly true for the brain, for which surgery carries high risks, and ionizing radiation can be damaging in even lower quantities. The use of fmri scans to obtain time course data of an entire brain volume is one of the first opportunities researchers have had to correlate physiological and psychological behavior with biological processes in the brain. When the BOLD contrast imaging technique was shown to correlate with activation of individual brain regions, research began into mapping the brain, a field which is still only understood at its most basic levels. fmri has become the sole modality used for brain mapping since it was introduced in the early 1990s. Initial hurdles faced with fmri imaging included the low spatial resolution and signal to noise ratio of the rapidly obtained images. Such problems were marginally improved by fine tuning sequence parameters, but the biggest improvements lay with hardware innovation and the move to higher powered MRI machines. The Robarts research facility at the University of Western Ontario has, in addition to its more common 3T Siemens whole body scanner, a recently developed and acquired 7T human head scanner. This scanner was brought to the facility in 2009 and used for its first clinical study in 2011, The purpose of this research was to obtain equivalent image sets from both machines given the constraints of the hardware involved, and to compare, both qualitatively and quantitatively, the image sets from each machine. 1 P a g e

3 Theory The Physics of Magnetic Resonance Imaging MRI is an imaging modality reliant on the nuclear resonance properties of tissues. Every atom has a nuclear magnetic spin, an inherent property of atoms (as a consequence of it being a property of fundamental particles), which dictates its rotation about an axis through its centre. These magnetic spins give the atom a magnetic dipole, the fundamental principle exploited in MRI. Any magnetic dipole will align itself with a larger external magnetic field. Without an external magnetic field all the dipoles will be aligned randomly. This is where the primary B 0 field comes into play in MRI. The B0 field is the strongest field used in the MRI machine, and is the field strength given to the name of the machine, in our case 3T and 7T. This field is aligned with the z-axis of the machine, in line with the subject lying inside. The strength of this field causes all of the diploes in a subject to align with the z-axis as seen in figure 1, a starting point from which we can manipulate the dipoles. Note that the spins can align either parallel or anti-parallel to the external field. This becomes important when looking at signal to noise ratio later. Figure 1: Showing the magnetic dipoles of hydrogen atoms (protons) randomly aligned, and then aligned with an external magnetic field along its axis. The next step in MRI is to force the dipoles to precess around the z-axis they have been aligned to. First, an understanding of precession is required. Precession occurs when an object rotating on an axis experiences a torque in a direction other than that of the primary rotation axis. This causes the spin 2 P a g e

4 axis of the object to rotate itself around the primary rotation axis. In the case of the magnetic dipole of the atom, we can cause the precession of the spin axis around the axis of the B 0 field as seen in figure 2. Figure 2: Showing the spin axis of a hydrogen atom precessing about the primary axis of rotation, that of the B 0 field. The angular momentum of its precession can be determined with the following formula: ω 0 = γ B 0 = f 2 π Equation 1: for determining precession frequency Using this formula the angular frequency ω 0 (a measure of the rate of rotation, measured in radians per second) can be found from the gyromagnetic magnetic ratio γ (an inherent measure of the strength of the magnetic moment, unique for each atom, measured in MHz per Tesla) and the primary magnetic field strength B 0 (measured in Tesla). The frequency of precession (another measure of the rate of precession, measured in Hz) can also be found, as it is equivalent to the angular frequency divided by 2 π. As the gyromagnetic ratio for each atom is unique, so too is the angular frequency of precession. It is in this way that a particular type of atom can be selected for measurement in an MRI image. The frequency is the rate of precession about the B 0 axis for any given atom. If we create an RF pulse at the same frequency, we can cause all atoms of that type in a sample to tilt away from the B 0 axis and begin precessing. This phenomenon is known as resonance. This is what the second magnetic field, the B 1 field, is used for. An RF pulse is created by an RF coil rotating in the plane transverse to the B 0 field at the precession frequency. While it is much weaker than the B 0 field, it does rise proportionally to the B0 field, giving rise to some inhomogeneity which will be discussed later. While an atom is precessing it is in a state of imbalance. After the RF pulse has started the precession, the magnetization vector of the dipole will return to its equilibrium state. During the return 3 P a g e

5 to equilibrium, the atom emits its own RF pulse (an emission of the energy it required to put it into a precessing state). This energy is detected by the RF coil, and recorded as image data. Each tissue type emits a unique pulse and so can be differentiated from the surrounding tissue. For example, white and grey matter in the brain both contain ample water, and therefore hydrogen, but the energy emitted from the hydrogen atoms in each will be different. This information can tell us about the amount of an atom, and the tissue type it resides in, but does not include information on its location. This is where the third magnetic field used in MRI comes in; the Gradient fields. The gradient fields consist of three graded magnetic fields parallel to each axis, stronger at one end of the axis then the other. All three field gradients have the same properties, but are applied at distinct moments in different directions to spatially encode the entire image. The first field applied is along the z-axis. This field is the slice selection gradient (GSS) and alters the strength of the B 0 field along the z-axis just enough so that only a particular plane (perpendicular to the z-axis) is subjected to the exact precession frequency. The second field applied acts along the y-axis. This field is the phase encoding gradient (GPE) and acts for a short time to alter the phase of each row of atoms without affecting the frequency. In this way all of the atoms in the plane are still precessing, but each row is slightly phase shifted, which leads to the image signal being slightly out of phase, and encoding for position along the y-axis. The third field applied is along the x-axis. This field is the frequency encoding gradient (GFE) and acts to alter the receiving frequencies along the x-axis. In this way each column of atoms has shifted frequencies which encode their position along the x-axis. Every atoms position in space can be determined from the use of these three gradient fields. Figure 3 shows an image of the gradient coils and their respective field axes. 4 P a g e

6 Figure 3: Showing the three gradient coils and their respective axes, used to spatially encode the MRI signal data. With those basic principles of MRI determined, one can go about creating an MRI sequence, which includes the previous discussed parameters as well as a number of parameters I will not go into detail on as they are beyond the scope of this project including: Echo types and Contrast type (determined by varying T1 and T2 times), reconstruction methods, sequence acceleration, and artifact reduction. These variables determine a sequence within you can further alter another set of parameters. These include the TR (the time between two RF pulses), the TE (the time between the RF pulse and the signal data being collected), and the field of view (the sections of the entire MRI field for which signal will be recorded). The MRI sequence used in this experiment was a Magnetization Prepared Rapid Acquisition Gradient Echo (MPRAGE). This sequence is used to obtain a high contrast, high spatial resolution 3D structural. This can be used as a high quality reference image for the fmri data we obtain. This sequence is T1 weighted, giving us a stereotypical MRI image in which fats appear brighter than water. 5 P a g e

7 The Mechanics of fmri Functional magnetic resonance imaging (fmri) is a form of MRI adapted to measure brain activity with high temporal resolution. It relies on fast repetitive imaging sequences which collect entire brain volumes on a second by second basis. fmri uses BOLD contrast (blood oxygen level dependent) to detect active brain regions. BOLD Contrast relies on a sequence of physiological events following brain activation. First the brain is activated with a task (motor control, such as thumb movement is popular. A movie with auditory and visual stimulation was used in our experiment). O2 consumption to the activated regions of the brain is increased, and local blood flow increases within that region. The ratio of oxyhemoglobin to deoxyhemoglobin increases in the region due to the increased influx of oxygenated blood. This ratio increase is detected as a weak transient rise in a T2 weighted signal. Thus areas of the brain being activated at any particular point in time will show as having a higher signal (brighter on our image). The fmri sequence used in this experiment was an Echo Planar Imaging (EPI) sequence. This is a fast repetitive imaging sequence which provides us with an entire brain volume in a short period of time (TR = 2 seconds for this experiment). The trade-off for this rapid image acquisition is the decreased spatial frequency. This sequence is T2 weighted, giving a less traditional image in which water appears brighter than fats. High Powered MRI Increased B0 field strength in MRI has a number of theoretical benefits and consequences, a few of which will be discussed in this paper. One of the primary benefits of high powered MRI is the increased signal to noise ratio (SNR). From Boltzmann Distribution, an increase in field strength should 6 P a g e

8 accentuate the difference in parallel and anti-parallel spins, increasing the signal to noise ratio. The potential signal will vary with the square of the B0 field, while the noise will progress linearly. Thus the SNR should increase linearly with field strength. SNR is a measurement which can be very difficult to measure, especially in MRI images. SNR at its core it determined from the following relationship: Signal-to-Noise-Ratio = Mean Signal in a region / Mean Noise in a Region However determining true signal and noise measurements is nearly impossible. Signal measurements are always affected by noise, and vary across the image. Noise in the image comes from many sources. The noise we want to isolate is that due to the scanner collecting the data, the noise which is not representative of any biological structure or process. However, there are many sources of noise intermingled from physiological processes, natural noise in the brain activation, and noise from slight movements of the subject which we might later correct for. Most MRI SNR measurements simply use the average signal in a given region of interest (ROI) as the signal measurement, ignoring the proportionally tiny contribution of noise. Noise measurements are collected in various ways. A method used in a prior study of high field strength MRI used the mean of the artifact free image background (outside the skull at the edge of the image) as a way of determining the noise which could not be due to physiological processes. This paper uses a slightly more common method in imaging, which us to take the standard deviation of signal across the signal as the noise for the entire image. A similar alternative is to use the standard deviation from only the ROI the mean signal was found in. Relaxation times are also directly proportional to the B 0 field strength. Therefore TR times and overall scan times are to be expected to increase at high field strength. Also the specific absorption rate (SAR), which is a measure of the energy deposited into the body being scanned, increases with the square of the B 0 field. This means that certain sequences cannot be as long, or cannot be done back to 7 P a g e

9 back under health regulations. Another downfall is that the auditory levels increase in the higher powered scanner, giving it some potential to interfere with the auditory portion of our fmri stimuli. It was mentioned early that the increased B 1 field or RF pulse is subject to some inhomogeneity at higher field strength. This is due to a standing wave pattern. As B 0 increases, the precession frequency increases proportionally. Precession frequency is related to wavelength by the following formula: λ = v / f Equation 3: relating precession frequency to wavelength Using this formula the wavelength λ (the distance between peaks of the electromagnetic wave, in meters) is equal to the speed v (of the electromagnetic wave front, measured in meters per second, in this case v is the speed of light 3x10 8 m/s) divided by the precession frequency f (a measure of the rate of oscillation of the wave, measured in Hz). Thus, as precession frequency increases due to increased B 0 field strength, the wavelength of the RF pulse decreases. In low powered MRI machines the 3T machine, the wavelength is long enough that the portion of the wave inside the scanner bore is essentially linear. However, in high powered machines like the 7T machine the wavelength is short enough that the portion of the wave in the scanner bore has significant amplitude differences throughout. This is magnified by the use of modern RF coils, which have multiple coils and produce nodes of low signal throughout the image. Another source of inhomogeneity is due to the size of the head bore in the 7T machine. There is very little space inside the head bore, and the subjects head is very close to the RF coils. The nature of the RF coils means they create a less homogenous field near the edges of the head bore, giving the images a squared off look to the top and back of head, along with distortion in the face. 8 P a g e

10 Methods Data Acquisition and Image Analysis After the initial decision was made to compare the two machines at Robarts, appointments were made for both machines in back to back time slots, to help avoid variation in our subject or testing patterns. Two sets of images were obtained from each machine. From both machines an MPRAGE structural was obtained. This gives us a single high quality brain volume, which is an excellent reference image, and is good for qualitatively comparing the machines. Next two EPI sequences were obtained from each machine. The first EPI was a resting state image set, in which the subject was to lie still, with his eyes open or closed, but not asleep. This serves as a base line reference to compare the brain activation in the next EPI against. The first six volumes collected were dummy volumes, and are not used in later analysis. This negates any noise from the machine beginning its sequence and the subject possibly reacting to it. The second EPI was collected while the subject watched a movie. This movie was projected (front projection for both machines) for the subject to see, and audio was fed to the subject through headphones. The constant audio visual stimulation provides a large amount of activation to be studied. A few other scans, including an MP2RAGE and field maps, were obtained as well, but were not used in this experiment. Image data was delivered to the Cusack Lab imaging server in dicom format. The data from the 3T scans was delivered separated by sequence, with sequence parameters stored in the dicom header. The data from the 7T scans was delivered as one large block of dicom image files in the order the sequences were conducted. This is one particularly annoying disadvantage of the 7T scanner and software. 9 P a g e

11 Analysis of the images was done using Matlab Software Modules SPM (Statistical Parametric Mapping), a well developed, well documented, free Matlab module evolved from early MRI software used in the 90s. The second piece of software used was AA (Automatic Analysis), a piece of software developed in Cambridge by a small team including my supervisor Dr. Rhodri Cusack. AA is used first and automatically performs a number of tasks. The first and most relevant step is the conversion of dicom image sets into more easily analyzed NIfTI files. AA then goes on to run a number of postprocessing corrections on the image sets, such as image realignment and image smoothing, although for this experiment we worked primarily with the raw NIfTI files. With the files in NIfTI format they are easily manipulated with SPM which can read the images into Matlab as 3D or 4D (in the case of our time course EPIs) matrices. Spatial SNR Calculations The first section of the experiment was to determine the spatial SNR of the four sets of EPIs as a comparison of the two machines. As discussed in the theory section, SNR was calculated from the mean signal in an ROI and the standard deviation of the entire image. First an ROI was selected. We chose the auditory cortex due its previously established activity levels with our movie stimuli. Using SPM, a binary image of the auditory cortex was resliced to align it with our EPI sets (separately for the 3T and 7T image sets). It was then applied to each volume of the EPI sequence to isolate voxels in the ROI region. The signal mean of these voxels was calculated and divided by the standard deviation of all the voxels in the volume. This was done with a Matlab script I wrote from scratch just for the image set. The script created a matrix of mean signals, standard deviations, and SNR values for every volume fed into the script. The SNR values were exported to Excel for statistical analysis. 10 P a g e

12 Excel was used to compare the exported SNRs. Four sets of data were exported; 171 SNR values for both the 3R resting and 3T movie EPI sets, as well as 240 SNR values for both the 7T resting and 7T movie EPI sets. One-tailed paired t-tests were used to analyze the data. Individual Component Analysis The second part of the experiment was to run Individual Component Analysis (ICA) on the acquired data. ICA is a relatively new and incredibly value type of fmri analysis. Without any predefinition of important spatial regions or temporal profiles, ICA can take an entire fmri time course and determine brain activity patterns in either the spatial or temporal domain. For example, one of the reported components of ICA for our fmri set was for the auditory cortex. The ICA analysis recognizes that the signal values in that region are changing in sync with each other across the time course, and reports it as a component, although it does not automatically identify the brain region it belongs to. ICA returns a number of noise related components, due to the consistent nature of many sources of noise (an oscillating inhomogeneity, or consistent movement of the subject s head. The software we used for ICA analysis was FSL. Three sets of data were provided to FSL; the 3T EPI sequences (both resting and movie), the 7T EPI sequences (both resting and movie), and a group study done previously in Cambridge with similar parameters in a 3T MRI (Using the same movie stimuli). The Group Study, consisting of 20 subjects, serves as a type of average data set for comparing the component analysis of our data sets with. A summary of the data sets is shown in table P a g e

13 3T EPI 7T EPI 3T Group Study TR=2, ipat=3, 2x2x2 voxel size 240 volumes (includes 6 dummies) Last 12s of movie missing 38 slices, descending sequential Screen crash around volume 185 TR=2 240 volumes (includes 6 dummies) Last 12s of movie missing 48 slices, descending sequential TR=2.47, multi-echo (5) 193 volumes (no dummies) First 5 volumes discarded in analysis first 12.5s of movie missing 32 slices, descending sequential 20 subjects Table 1: Summarizing the parameters of the three data sets submitted for ICA analysis ICA was run twice with two slightly different sets of images. During acquisition of the 3T EPI movie sequence the projection screen fell onto the subject, compromising the last 60 volumes of the data. ICA was run once on the entire data set, which produces a number of excess noise components in the 3T data, and a second time omitting the last 60 volumes from all data sets (to keep the time course matched between sets). These sets were named Dummycut and Moviecut, respectively. A Summary of the parameters of each analysis is shown in table 2. ICA - Moviecut ICA - Dummycut Uses raw data, pre-processed with FSL Excludes last 60 volumes (to avoid screen crash) Excludes 6 dummy scans 173 volumes in total For comparison to group ICA results, the first 12s of the movie need to be cut because group ICA discards first 5 volumes Need to adjust movie soundtrack to correlate with auditory component timecourse Uses raw data, pre-processed with FSL Excludes 6 dummy scans 236 volumes in total For comparison to group ICA results, the first 12s of the movie need to be cut because group ICA discards first 5 volumes Need to adjust movie soundtrack to correlate with auditory component timecourse Table 2: Summarizing the distinctions between the two ICA analyses used 12 P a g e

14 FSL processes the data and returns a list of all suspected components, with images of the component brain regions highlighted in a number of slices through a brain volume. All components were correlated with the known and identified components from the group study. This allowed us to identify the component representing the auditory cortex. The components representing the auditory cortex in all three sets of data were then correlated against each other, and against the sound envelope from the stimulus movie. The sound envelope is a simple time course of the intensity of the audio track in the movie, the theory being that peaks in audio intensity should correspond to peaks in activation of the auditory cortex. Results Sample Images For reference and qualitative comparison four sample images have been provided below. The two MPRAGE structurals and a sample volume of the EPI sequences are shown below. Figure 3: 3T MPRAGE Structural Image 13 P a g e

15 Figure 4: 7T MPRAGE Structural Figure 5: 3T EPI fmri 14 P a g e

16 Figure 6: 7T EPI fmri A quick comparison of the 3T and 7T images clearly shows the improved SNR, with much of the grainy background noise not existent in the 7T images. Also apparent is the previously discussed field inhomogeneity, expressed in the squared off top and back of the head visible in figure 4. You can also see some darker patches in figure 4 towards the top and each side of the brain, caused by the standing field patterns. Spatial SNR Calculations Spatial SNR for each volume was found using the custom Matlab script. The mean SNR found for the four EPI image sets are summarized table 3. 3T EPI - 3T EPI - 7T EPI - 7T EPI - Resting Movie Resting Movie SNR considering mean signal in auditory cortex ROI and using standard deviation of image as noise (average of 5 volumes) Table 3: Showing average SNR for each of the four EPI image sets 15 P a g e

17 Excel was used to conduct One-Tailed Paired Two Sample t-tests between the two 3T scans, the two 7T scans, the two resting state scans, and the two movie scans. Tables 4 7 show the results of the t-tests. t-test: Paired Two Sample for Means 3T resting and 3T movie 3T resting 3T movie Mean Variance 1.98E E-05 Observations (n) P(T<=t) one-tail 9.26E-08 t Critical one-tail Table 4: The results of the t-test comparing the 3T resting and 3T movie SNR values. The result of the test indicates a significant difference between the 3T resting and 3T movie SNR values. t-test: Paired Two Sample for Means 7T resting and 7T movie 7T resting 7T movie Mean Variance E-05 Observations (n) P(T<=t) one-tail 1.4E-35 t Critical one-tail Table 5: The results of the t-test comparing the 7T resting and 7T movie SNR values. The result of the test indicates a significant difference between the 7T resting and 7T movie SNR values. t-test: Paired Two Sample for Means 3T resting and 7T resting 3T resting 7T resting Mean Variance 1.98E E-05 Observations (n) P(T<=t) one-tail 8.6E-240 t Critical one-tail Table 6: The results of the t-test comparing the 3T resting and 7T resting SNR values. The result of the test indicates a significant difference between the 3T resting and 3T resting SNR values. 16 P a g e

18 t-test: Paired Two Sample for Means 3T movie and 7T movie 3T movie 7T movie Mean Variance (n) 3.14E E-05 Observations P(T<=t) one-tail 3.6E-245 t Critical one-tail Table 7: The results of the t-test comparing the 3T movie and 7T movie SNR values. The result of the test indicates a significant difference between the 3T movie and 7T movie SNR values. The results of all four t-tests indicated a significant difference between the relevant sets of spatial SNRs. These results validate two key points. First, the tests between the two 3T scans and the two 7T scans demonstrate the signal increase due to brain activation when the subject experiences the movie stimuli. As discussed in the theory section this increase is very small, an increase of 0.16% in the 3T data and 0.88% in the 7T data. The second, and more relevant, point the test demonstrate is the increased SNR in the 7T scanner compared to that of the 3T scanner. A 16.0% increase is seen in SNR values of the resting states between the two machines, while a 17.4% increase is seen in SNR values of the stimulated states between the two machines. ICA Results The first set of results returned from FSL includes a summary of all found components and their corresponding images. The non-noise components were manually identified. This was done for both the Moviecut and Dummycut analyses. The summary of Components for the Moviecut and Dummycut analyses, respectively, are shown below in tables 8 and P a g e

19 3T EPI 7T EPI 3T Group Study Total components: 17 Non-noise components: 4 Ratio: 0.24 Non-noise ID: 11, 12, 13, 15 Total components: 32 Non-noise components: 12 Ratio: 0.38 Non-noise ID: 7, 10, 15, 19, 21, 22, 23, 25, 26, 27, 29, 32 Total components: 57 Non-noise components: 18 Ratio: 0.32 Non-noise ID: 1-11, 13, 17, 20, 23, 30, 31, 54 Table 8: Summarizing the component findings of the Moviecut ICA analysis. 3T EPI 7T EPI 3T Group Study Total components: 56 Non-noise components: 8 Ratio: 0.14 Non-noise ID: 37, 40, 47, 48, 51, 52, 53 Total components: 31 Non-noise components: 9 Ratio: 0.29 Non-noise ID: 7, 14, 16, 19, 22, 23, 24, 26, 30 Total components: 57 Non-noise components: 18 Ratio: 0.32 Non-noise ID: 1-11, 13, 17, 20, 23, 30, 31, 54 Table 9: Summarizing the component findings of the Dummycut ICA analysis. After the non-noise components were identified, correlation was run comparing the components from both our 3T and 7T data with that of the group study. Previous work with the group study data has identified non-noise components related to specific brain regions. By running this type of correlation we can easily determine which of the components in our 3T and 7T data correspond to particular brain regions. In our case, we were trying to identify the component representing the auditory cortex. Correlation was run with a significance p<0.05. The results of the correlation are summarized below in tables 10 and P a g e

20 3T EPI 3T Group Study 7T EPI 3T Group Study Group 3T Study total: 5 pairs: 1-9, 11-2, 13-3, 15-23, 17-9 Auditory: 11 total: 7 pairs: 1-16, 6-17, 7-2, 10-9, 17-16, 22-4, Auditory: 7 Known non-noise components (18): 1 = Visual 2 = Auditory 3 = Frontal-Parietal 4 = Frontal 5-11, 13, 17, 20, 23, 30, 31, 54 = other Table 10: Summarizing the correlation of the components of our 3T and 7T data with that of the known 3T group study data for the Moviecut ICA analysis. 3T EPI 3T Group Study 7T EPI 3T Group Study Group 3T Study total: 25 pairs: 1-11, 2-16, 7-5, 9-6, 13-6, 14-8, 16-1, 17-3, 20-31, 21-9, 25-8, 27-9, 30-16, 32-13, 36-20, 37-2 Auditory: 37 total: 15 pairs: 2-17, 6-13, 7-2, 9-17, 11-1, 12-8, 15-6, 20-9, 21-16, 22-4, 23-5, 24-23, 26-6, 29-13, 30-8 Auditory: 7 Known non-noise components (18): 1 = Visual 2 = Auditory 3 = Frontal-Parietal 4 = Frontal 5-11, 13, 17, 20, 23, 30, 31, 54 = other Table 11: Summarizing the correlation of the components of our 3T and 7T data with that of the known 3T group study data for the Dummycut ICA analysis. The above correlation data shows that components 11, 7, 37, and 7 for their respective data sets represent the auditory cortex, as determined by their correlation with the know auditory component of the 3T group study analysis. Shown below in figure 7 is the image provided for component 2 of the 3T EPI in the Moviecut analysis. The corresponding image for the Dummycut analysis looks similar. 19 P a g e

21 Figure 7: The highlighted brain regions for component 2 of the 3T EPI Moviecut analysis. The highlighted brain regions correspond with what is classically thought of as the auditory cortex, confirming the correlations results. With the auditory cortex components isolated, correlation can now be done between our 3T and 7T auditory components and that of the 3T group study, as well as correlation between our 3T and 7T auditory components and the sound envelope produced from the audio track of the stimulus movie. Shown below in figure 8 are the normalized time courses for the Sound Envelope, 3T auditory component signal, and 7T auditory component signal used in the Moviecut analysis. Corresponding time courses for the Dummycut analysis look similar. 20 P a g e

22 Figure 8: Showing the normalized time courses for the Sound Envelope, 3T auditory component signal, and 7T auditory component signal used in the Moviecut analysis. Statistical correlations were run between the above time courses. Correlations were calculated between both our 3T and 7T data with the 3T group study data, as well as between both our 3T and 7T data with the sound envelope. Correlations were run for both the Moviecut and the Dummycut analyses. The results are summarized in tables 12 and 13 below. ICA Moviecut 3T 7T 3T Group 7T Group 3T Sound Envelope 7T Sound Envelope Correlations for auditory component r =.73, p<.0001 r = 0.82, p<.0001 r = 0.87, p<.0001 r =.33, p<.0001 r =.39, p<.0001 Table 12: Summarizing the correlations between the data sets of the Moviecut analysis. 21 P a g e

23 ICA Dummycut 3T 7T 3T Group 7T Group 3T Sound Envelope 7T Sound Envelope Correlations for auditory component r =.71, p<.0001 r = 0.82, p<.0001 r = 0.90, p<.0001 r =.23, p<.0005 r =.30, p<.0001 Table 13: Summarizing the correlations between the data sets of the Dummycut analysis. Note that unlike all other results shown, correlation of our 3T data with the sound envelope is shown at a significance level of as opposed to The above data shows strong correlations amongst the time courses. The correlations are stronger in the Moviecut data, due to the omission of the noise creating portion of the 3T EPI data set. Especially relevant to this experiment, note that in all four sets of correlations the 7T timecourse correlated more strongly with the true data sets (the sound envelope representing the system input, and the group study representing a corrected average). Discussion Previous papers (1) comparing low and high field strength MRIs have shown consistently increased SNR for the high field strength machines, in various regions of the brain. Our goal was to show that a similar relationship existed between the two MRI scanners at Robarts. Our results showed the results we expected. SNR was increased significantly for the 7T scanner. Interestingly the increase in SNR that was found for the Robarts machines was less than in paper (1). There are a number of possible reasons for this. The other paper calculated SNR values for five different regions of the brain, and used a different method of calculation, considering the mean of a background area their measure of noise. This 22 P a g e

24 in itself could easily explain the differences, although does not show that either method is better than the other, simply that they likely cannot be compared directly with each other, since the noise measurement is quite arbitrary. It is also possible that the auditory cortex ROI was in a region of nodal inhomogeneity, with a slightly lower average signal than other brain regions. Further tests using additional ROIs, and more exploration into the degree and possible correction of the field inhomogeneities would answer these questions. An interesting thing I came across when determining the SNRs for our data was related to the above noted choice of noise measurement. I also wrote a script to determine SNR using the mean of a background region as the noise measurement, as in paper (1). My results appeared to lack a pattern. Although I only ran it on a few images, the 3T images had higher SNR values, and there was the resting and movie states didn t seem to have any consistent effect. It is noted in paper (1) that the background region was artifact free and some post processing was mentioned. So it is possible that working with the raw image data, as we did, prevented us from accurately using this measurement. The ICA results were very interesting, and perhaps more relevant to future studies. ICA being a technique that is becoming more common for data analysis in research experiments, it bodes well for the 7T machine to see increased correlations in all comparisons with the 3T data. Correlation of additional ICA components would help to solidify the degree of improvement the 7T machine shows over the 3T. Although our numerical results show improved performance across the board for the 7T scanner, it was not without its drawbacks. Qualitative comparison of the resulting images shows that the field inhomogeneities are a large factor in the 7T scanner. This renders the 7T scanner almost useless for research of certain brain regions. Further study into correcting for this in either post-processing or by tweaking the sequence parameters, could potentially make the 7T scanner a universally better choice for fmri experiments. 23 P a g e

25 Conclusions The purpose of this research was to obtain equivalent image sets from both machines given the constraints of the hardware involved, and to compare, both qualitatively and quantitatively, the image sets from each machine. We found that the images we obtained from the 7T machine appeared to have noticeably higher SNRs than those of the 3T machine. However, field inhomogeneities were evident as both squared edges of the head in the image, and dark spots in certain areas of the brain. These inhomogeneities were not present in the 3T data. We found in our research that the 7T MRI provided a statistically significant increase in SNR over the 3T MRI, both with the subject in a resting state, and when presented stimulus; increases of 16.0% and 17.4% were observed respectively. We also validated an increased signal when a subject was presented stimulus in comparison with a resting state for both our 3T and 7T image sets; increases of 0.16% and 0.88% were observed respectively. ICA analysis showed good correlation of the auditory components of both our 3T and 7T data with previously analysed 3T group study data, and with the sound envelope of our auditory stimulus. In all comparable correlations the 7T data showed a higher coefficient of correlation than the 3T data. 24 P a g e

26 References 1. Vaughan J, Garwood M, Collins G, Liu W, DelaBarre L, Ardriany G, Andersen P, Merkle H, Goebel R, Smith M, Ugurbil K. 7T vs. 4T: RF Power, Homogeneity, and Signal-to-Noise Comparison in Head Images. Magnetic Resonance in Medicine 46: 24-30, Smith S, Fox P, Miller K, Glahn D, Fox P, Mackay C, Filippini N, Watkins K, Toro R, Laird A, Beckmann C, Raichle M. Correspondence of the Brain's Functional Architecture during Activation and Rest. Proceedings of the National Academy of Sciences of the United States of America 106: , Gelman N. Medical Biophysics 3505F Mathematical Transform Applications in Medical Biophysics. Lecture Slides, P a g e