WO2019028551A1

WO2019028551A1 - Systems and methods for assessing functional magnetic resonance imaging data quality using regularization of frequency-domain entrophy

Info

Publication number: WO2019028551A1
Application number: PCT/CA2018/050959
Authority: WO
Inventors: Steven BEYEA; Christopher O'GRADY; Steve Patterson
Original assignee: Nova Scotia Health Authority
Priority date: 2017-08-08
Filing date: 2018-08-07
Publication date: 2019-02-14

Abstract

Systems and methods are provided for performing functional magnetic resonance imaging (fMRI) and generating one or more fMRI data quality metrics. The one or more fMRI data quality metrics are generated based on the calculation of regularized frequency-domain entropy. For example, Tikhonov regularization may be employed for the generation of a regularized power spectrum, and spectral domain entropy may be calculated by processing the regularized power spectrum. The resulting regularized frequency-domain entropy values associated with fMRI voxels may be processed to generate one or more fMRI data quality metrics. For example, fMRI data quality metrics may be generated based on statistical measures associated with the regularized frequency domain entropy measures of the fMRI voxels. The fMRI data quality measures may be employed to provide a preliminary assessment of fMRI data quality, and may be employed to determine whether to perform additional fMRI scanning to obtain fMRI data of sufficient quality.

Description

SYSTEMS AND METHODS FOR ASSESSING FUNCTIONAL MAGNETIC RESONANCE IMAGING DATA QUALITY USING RE G U LARIZ ATI O N OF FREQUENCY-DOMAIN ENTROPY CROSS-REFERENCE TO RELATED APPLICATION

This application claims priority to U.S. Provisional Application No. 62/542,673, titled "SYSTEMS AND METHODS FOR ASSESSING

FUNCTIONAL MAGNETIC RESONANCE IMAGING DATA QUALITY USING REGULARIZATION OF FREQUENCY-DOMAIN ENTROPY" and filed on August 8, 2017, the entire contents of which is incorporated herein by reference.

BACKGROUND

The present disclosure relates to magnetic resonance imaging (MRI) and functional magnetic resonance imaging (fMRI).

fMRI involves the non-invasive imaging of brain function based on the detection of changes in blood flow via magnetic resonance imaging. fMRI image data quality may vary due to multiple reasons; e.g., patient movement and physiological effects such as respiration and cardiac pulsation/flow, differences in environmental noise, non-adherence by the patient to a functional task, and scanner drift and electronic noise. It may take hours or even days to fully process fMRI data, and waiting until the fMRI data is fully processed to assess data quality can be problematic. For example, if the patient has left, it may not be possible to re-scan the patient should the data quality be low. The rapid detection of the quality of fMRI data is therefore beneficial for determining the potential interpretability of the scan and the overall clinical utility of fMRI.

Existing methods for automated/real-time detection of fMRI data quality use proxy metrics, such as motion parameters. Methods such as the General Linear Model (GLM), first published by Karl Friston in 1994, detect the presence of useful task-related information using parametric modelling; a statistic is produced for each voxel based upon the beta values of the fit to an idealized task response ["Statistical parametric maps in functional imaging: a general linear approach. Human brain mapping" (1994) 2(4), 189-210]. While methods such as the GLM are sensitive to task information, they require significant processing steps and model assumptions. Model-free methods such as independent component analysis (ICA) detect signal components in the brain, but are no more sensitive to task-related data than noise or other unrelated signal sources. Accordingly, such methods may result in the identification of many signal components that may not relate to the desired information.

SUMMARY

Accordingly, in one aspect, there is provided a method of performing functional magnetic resonance imaging (fMRI) data quality assessment, the method comprising:

a) obtaining a set of fMRI training datasets having a variation in fMRI data quality;

b) processing the set of fMRI training datasets to perform frequency- domain regularization according to a range of regularization strengths;

c) selecting a regularization strength according to pre-selected performance criteria;

d) controlling a magnetic resonance imaging scanner to generate a sequence of RF pulses and detect RF signals that are responsively emitted by a subject during the performance of a series of tasks characterized by a task frequency, and processing the signals to generate a patient fMRI dataset; e) employing the selected regularization strength to perform

frequency-domain regularization of the power spectra of the patient fMRI dataset, wherein the regularization is dependent on the task frequency, thereby obtaining regularized power spectra; f) generating frequency-domain entropy measures from the

regularized power spectra, thereby obtaining a set of regularized frequency- domain entropy measures; and

g) generating one or more fMRI data quality metrics based on the set of regularized frequency-domain entropy measures.

In another aspect, there is provided a magnetic resonance imaging system for performing functional magnetic resonance imaging (fMRI), the system comprising:

a magnetic resonance imaging scanner; and

control and processing hardware operatively coupled to said magnetic resonance imaging scanner, wherein said control and processing hardware comprises memory coupled with one or more processors to store instructions, which when executed by the one or more processors, causes the one or more processors to perform operations comprising:

d) controlling the magnetic resonance imaging scanner to generate a sequence of RF pulses and detect RF signals that are responsively emitted by a subject during the performance of a series of tasks characterized by a task frequency, and processing the signals to generate a patient fMRI dataset; e) employing the selected regularization strength to perform frequency-domain regularization of the power spectra of the patient fMRI dataset, wherein the regularization is dependent on the task frequency, thereby obtaining regularized power spectra;

f) generating frequency-domain entropy measures from the

A further understanding of the functional and advantageous aspects of the disclosure can be realized by reference to the following detailed description and drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments will now be described, by way of example only, with reference to the drawings, in which:

FIGS. 1A and 1 B show (A) a "strong" fMRI signal showing clear response to a task, and (B) the power spectrum for the signal (shown in blue), and the multiplication factor for a (non-regularized) spectral entropy calculation (shown in red). Despite the strong peak at 0.02Hz, a relatively high spectral entropy value is obtained because of the noise at the other frequencies.

FIGS 2A-C are plots demonstrating how the effect of regularization increases with the difference of the power of the task frequency relative to the power of all other frequencies. Original power spectra are shown in blue, regularized power spectra are in red. In FIG. 2A, the task frequency power is comparable to the other frequencies' powers and the regularization has little effect. In FIG. 2B and 2C, the task frequency dominates the power spectra and the regularization exaggerates this effect, with FIG. 2C exhibiting a larger difference than FIG. 2B. The larger the difference between the power of the task frequency and other frequencies, the greater the regularization as evidenced by comparing FIG 2B and 2C.

FIG. 3A plots the distance to the top left corner of an ROC curve of spectral entropy in simulations. Low spectral entropy values were taken to be "active", and the optimal cutoff was chosen based on sensitivity and specificity. The results are plotted here as a function of regularization strength and percent signal change, and have been resampled to higher resolution using spline interpolation.

FIG. 3B plots the error for the FIG. 3A, calculated using 5 sets of simulations with the same levels of activation and the same ROI. Low error compared to actual values gives confidence.

FIGS. 4A and 4B plot the change in spectral entropy relationship to t- statistic with (FIG. 4A) and without (FIG. 4B) regularization. Values were taken from the active ROI only of simulations with percent signal change between 2-5%. Addition of regularization dramatically reduces the spectral entropy values within the active ROI.

FIGS. 5A-C demonstrate how spectral entropy changes much more dramatically with regularization strength for a signal containing task information than a signal that is purely noise. In FIG. 5A, the noisy signal's spectral entropy is relatively invariant, which is desired, but the task signal changes considerably. This demonstrates the data-driven nature of the regularization method. The task signal is shown in FIG. 5B, and the noisy signal is shown in FIG. 5C.

FIG. 6 is a flow chart for implementing regularized spectral entropy algorithm on MRI scanner as a data quality metric.

FIG. 7 is a block diagram of an example system for performing functional magnetic resonance imaging and for the automated determination of fMRI data quality.

FIG. 8 is a plot demonstrating how regularized spectral entropy is seen to decrease with increasing percent signal change. Note the linear relationship between regularized spectral entropy and percent signal change.

FIG. 9 is a plot demonstrating how regularized spectral entropy decreases with increasing t-statistic in the active ROI of simulations of varying percent signal change. Note the approximately linear relationship between regularized spectral entropy and conventional t-statistic.

FIG. 10 shows: (rightmost three images) axial, sagittal, and coronal views of a t-statistic map of the reduction in sample entropy from task to rest in 32 simulated data sets; (middle column of images) regularized entropy maps for three example simulations of percent signal change 2%, 3.7% and 5%; and (leftmost column of images) parametric maps of the same datasets as spectral entropy showing the "gold standard" in identification of task-active regions. Regularized spectral entropy, generated using the exact same conventional single-run data set as the GLM t-stat map, has better agreement to t-stat, while maintaining a sensitivity to the variations in data quality.

FIG. 11 plots the sensitivity and specificity of parametric maps (GLM) and regularized spectral entropy as a function of percent signal change. Error bars are from 5 replications of each simulated data set. Spectral entropy had sensitivity of 0.7296 +/- 0.0086, and specificity of 0.7532 +/- 0.005 for 32 simulations of varying percent signal change. All data determined using the simulated data sets described previously. While the GLM, as expected, has better sensitivity and specificity, it is also important to note that the GLM performs better in terms of sensitivity and specificity, but it also requires more assumptions and a priori knowledge, computation time, and pre-processing. Additionally, the GLM is intended to produce clear visual maps but not in-and- of-itself provide a metric of scan quality.

FIG. 12 shows an example of an fMRI language scan from a representative subject. Regions in yellow have low regularized spectral entropy (below 0.2), regions in blue have high t-statistic (above 7) found using the GLM, and regions in green are overlap of low regularized spectral entropy and high t-statistic. A high degree of overlap occurs in a language task activated region in the left cortex.

FIGS. 13A and 13B show t-statistic histograms with voxels grouped based on regularized spectral entropy values. High values of regularized spectral entropy are shown in FIG. 13A, and low regularized spectral entropy is shown in FIG. 13B. Low values of regularized spectral entropy favors more extreme values of t-statistic, indicating a relationship to the amount of task- related information content.

FIG. 14 shows the difference in regularized spectral entropy distributions on the same subject when performing the task (blue) and resting (red). The inset shows the greater fraction of low regularized spectral entropy voxels when the subject is actually performing the task. FIG. 15 plots the average regularized spectral entropy distributions of four language scans. Error between values is shown in yellow, and the average value in the dark line. This demonstration consistency of regularized spectral entropy distributions, and replicability between subjects.

DETAILED DESCRIPTION

Various embodiments and aspects of the disclosure will be described with reference to details discussed below. The following description and drawings are illustrative of the disclosure and are not to be construed as limiting the disclosure. Numerous specific details are described to provide a thorough understanding of various embodiments of the present disclosure. However, in certain instances, well-known or conventional details are not described in order to provide a concise discussion of embodiments of the present disclosure.

As used herein, the terms "comprises" and "comprising" are to be construed as being inclusive and open ended, and not exclusive. Specifically, when used in the specification and claims, the terms "comprises" and

"comprising" and variations thereof mean the specified features, steps or components are included. These terms are not to be interpreted to exclude the presence of other features, steps or components.

As used herein, the term "exemplary" means "serving as an example, instance, or illustration," and should not be construed as preferred or advantageous over other configurations disclosed herein.

As used herein, the terms "about" and "approximately" are meant to cover variations that may exist in the upper and lower limits of the ranges of values, such as variations in properties, parameters, and dimensions. Unless otherwise specified, the terms "about" and "approximately" mean plus or minus 25 percent or less.

It is to be understood that unless otherwise specified, any specified range or group is as a shorthand way of referring to each and every member of a range or group individually, as well as each and every possible sub-range or sub -group encompassed therein and similarly with respect to any subranges or sub-groups therein. Unless otherwise specified, the present disclosure relates to and explicitly incorporates each and every specific member and combination of sub-ranges or sub-groups.

As used herein, the term "on the order of", when used in conjunction with a quantity or parameter, refers to a range spanning approximately one tenth to ten times the stated quantity or parameter.

The present disclosure provides systems and methods for the rapid assessment of fMRI data quality. In some example embodiments, a data- driven metric for functional Magnetic Resonance Imaging (fMRI) data quality is provided that is automated and objective. In some example

implementations, such a metric may be implemented directly on an MRI scanner and used to provide direct feedback on the information content quality of the data acquired at the time of the scan. The example systems and methods disclosed herein for objective fMRI data quality characterization can be implemented on fMRI scanners to automatically provide quality assurance both during and after the fMRI scan.

Various example embodiments of the present disclosure employ the application of information theory to functional MRI (fMRI), motivated by the goal of detecting useful task information in raw signals. The reasoning behind use of raw or minimally-processed signals is to rapidly detect task-relevant information immediately after the scan, or ideally during the scan itself.

Additionally, minimal processing prevents concealing artifacts that reduce functional map quality, even if the reduction in quality is non-obvious after processing (for example, missing or reduced activation). As described in detail below, various example embodiments of the present disclosure employ spectral entropy in combination with a data-driven frequency-domain regularization scheme to form a method that facilitates the rapid identification of the presence of useful information in raw fMRI data. Unlike currently available methods of generating fMRI data quality metrics, the example embodiments disclosed herein may provide, in either real-time or immediately following a scan, information on the fundamental relevant information content of the data that was collected.

Existing methods for automated/real-time detection of fMRI data quality use proxy metrics, such as motion parameters or SNR measurements, and do not take a holistic approach that incorporates all sources of noise or non-task based signal vs. task-based signal.

Every fMRI scan contains a variety of noise sources that are inhibitory for identification of brain regions responding to a task. These noise sources may come from the scanner itself, such as digitization error, magnetic drift, etc., or they may be inherent to the patient themselves. Patient sourced noise includes physiological phenomena like breathing or heart rate. In contrast, according to various example embodiments of the present disclosure, information theory may be employed to facilitate the generation of fMRI quality metrics that involve the detection of task-related data on raw, minimally processing signals with little prior information. Indeed, in some example embodiments, the only prior information needed to generate one or more fMRI quality metrics is the task frequency. In contrast, conventional GLM fMRI analysis requires an assumed hemodynamic response function as an input model, which is convolved with the task design, knowledge of the relative slice timings, etc.

Data entropy, a form of information theory, may be described as a measure of the average information content in a signal. Many forms of entropy calculation exist, including Shannon entropy, sample entropy, spectral entropy, and mutual information. Some of these methods have been previously applied in the limited context of analyzing fMRI data to examine differences in entropy between individuals due to brain disorders, between brain regions because of tissue differences, and within brains because of task versus resting state. One study by Alpert et al. ["Temporal characteristics of audiovisual information processing" (2008) Journal of neuroscience, 28(20), 5344-5349] analyzed the latency in information processing by computing the time at which mutual information was maximized between signals and stimulus through the brain.

Other work in applying information theory to fMRI has attempted to investigate whether or not additional information is gained from simultaneous EEC Measures such as mutual information, a function of entropy, are particularly useful for such applications that involve the detection of similarity of signals or transfer of information.

Entropy, on its own, is suited to quantifying information content. A study by Wang et al. ["Brain entropy mapping using fMRI" (2014) PLoS ONE, 9(3)] demonstrated a typical distribution of sample entropy in the brain, and interestingly, also a reduction in sample entropy in task activated regions. This is complemented by the work of De Araujo et al. ["Shannon entropy applied to the analysis of event-related fMRI time series" (2003) Neurolmage, 20(1 ), 31 1-317] which used Shannon entropy to identify task activation in event- related paradigms.

Although a wide range of applications exists for the application of information theory in fMRI, few applications focus on isolating and identifying information that is task-illuminative: data that is informative regarding whether or not the task is being completed and how much of the signal can be attributed to that task. Indeed, previous implementations of entropy analysis in fMRI were only able to exhibit entropy changes when comparing multiple data sets. Moreover, none of the previous implementations of entropy analysis demonstrated a capability of generating measures correlated with variations in data quality. The present inventors set out the develop an entropy-associated method that would be capable of processing raw fMRI data in order to provide one or more measures associated with the quality of the fMRI data. It was deemed that the following properties or capabilities would be desirable and beneficial for a suitable measure of fMRI quality: a quality metric that varies with changes in functional data quality; a quality metric having robust sensitivity over a realistic data quality range; a quality metric that may be calculated using a single standard block-design paradigm that is commonly used in fMRI (including for the vast majority of clinical fMRI tests of language, cognitions, and motor/sensory mapping); and a quality metric that may be calculable with computation times on the order of seconds or minutes using standard computers.

The present inventors have found that the generation of fMRI quality metrics that are sensitive to data quality variation can be achieved using a regularization ensuring that the entropy is responsive to a specific type of information content (i.e. responsive to information content specific to task- related information in the fMRI signal.). Various example embodiments of the present disclosure employ a regularized entropy measure, determined in the frequency domain, (e.g. Shannon entropy in the frequency domain) as a method of identification of task-related functional brain information. In some example embodiments, the regularized frequency-domain entropy is employed to calculate one or more metrics of fMRI data quality. In particular, in some example embodiments, Tikhonov regularized spectral entropy is calculated for fMRI block-design time-courses in the brain. Regularization is employed to counteract the noise in fMRI that is always present, even in the most robust task response.

According to some example embodiments, regularization and entropy calculation are performed in the frequency domain with regularization parameters that are selected using sensitivity and specificity for useful task signals. The regularization of entropy is designed to make a subjectively sparse power spectra objectively sparse according to spectral entropy. In this context, a "sparse" signal is one in which most of the true information contained within the signal is confined to very few frequency components. This is a known expected behavior for signals acquired from block-designed fMRI paradigms, as the resulting signal is essentially a periodic boxcar oscillation. For a signal where it is clear that the task frequency is dominating the signal, a properly regularized spectral entropy calculation will accentuate this in such a way that entropy will decrease and reflect the fact that the signal is sparse in overall frequency information and rich in task-related information.

A signal that is dominated by a specific frequency - e.g., the task frequency - will exhibit sparse power spectra in the sense that a majority of the information content is associated with only a single Fourier coefficient. This represents a low entropy situation. A noisy signal, on the other hand, with many contributing frequencies of comparable magnitude, will not be information-sparse and will therefore have high entropy in the frequency domain. According the example methods described herein, regularization may be employed to help increase the effective sparsity of a task-dominated signal, such that low entropy voxels will correspond to high t-statistic in the GLM as well as high SNR. A closely related finding, as discussed below, is that regularized spectral entropy spatially conforms to areas of known activation. The replicability of the present methods is also demonstrated by showing that an innate reliable distribution of regularized spectral entropy exists across subjects for a given scan type.

In various examples provided in the present disclosure, the spectral entropy of the time-courses from each voxel in the brain are determined, and characterization of the regularized frequency-domain entropy values is shown to reflect change in the raw fMRI data quality. This regularized frequency- domain entropy is shown below to be directly related to other measures of fMRI data quality (e.g., functional contrast sensitivity as measured by percent signal change). One aspect of the example methods provided herein is the data-driven method by which the regulanzation is chosen. In some example

embodiments, the regulanzation strength is determined (e.g. optimized) in order to generate one or more fMRI quality metrics that are sensitive to variations in data quality. If the fMRI data is regularized too aggressively, it will always return the a priori expected result and will not reflect the true measurement, however there is no benefit to using a regularization scheme if the regularization is too mild. Accordingly, various example embodiments of the present disclosure employ a data-driven method to achieve the selection of a regularization that facilitates (e.g. in some example implementation, optimizes) the calculation of spectral entropy in a manner that provides a suitable measure of fMRI data quality.

In some example embodiments, a data-driven regularized (e.g.

Tikhonov regularized) spectral entropy calculation method is employed to provide a measure of fMRI data quality (one or more fMRI data quality metrics), where regularization strength (e.g. a regularization parameter) is determined (e.g. objectively optimized) on the basis of a cost-function, such as based on maximizing sensitivity and specificity of training data sets. The regularization strength may thus be determined based on the processing of a plurality of fMRI datasets (e.g. training datasets) in which data quality is varied. The selected regularization strength (parameter) may then be applied during the spectral entropy calculation of one or more new (additional) fMRI data sets to provide a measure of the data quality of the new (additional) data sets. For example, the regularized frequency-domain entropy may be calculated for each voxel's time course, following volume realignment and low-pass filtering to remove the effect of scanner drift, with the resulting entropy values used to provide a metric for quality assessment.

Accordingly, in some example embodiments, regularization is employed prior to calculating spectral entropy, with a data-driven

determination of the regularization strength, in order to achieve sensitivity to variations in fMRI data quality. The present example methods involving regularized spectral entropy have been found to have comparable sensitivity and specificity to the General Linear Model (GLM), as well as a correlated response to percent signal change and t-statistic. Additionally, the present example methods involving Spectral Entropy have been found to be computationally fast and required less (e.g. minimal) a-priori information compared to other methods used to identify useful task-related information.

Referring now to FIG. 1A, a simulated signal from an example fMRI time course with good data quality is plotted, and the associated power spectrum is plotted in FIG. 1 B. FIG. 1 B also shows the multiplication factor for a (non-regularized) spectral entropy calculation . The signal and power spectrum correspond to a single voxel. The signal shown in FIG. 1 B is a sparse signal that favors the task frequency, and the power spectrum demonstrates that a majority of the total power resides at a frequency corresponding to the frequency of the task (i.e., the frequency of the active versus control blocks). However, the spectral entropy value is also affected by the presence of noise peaks. Even in the presence of the strong peak at 0.02 Hz, the power spectrum shown in FIG. 1 B will result in a relatively high spectral entropy value because of the many component noise frequencies and the normalization involved in the spectral entropy calculation. In some example embodiments, Tikhonov regulanzation (for an ill-fitted problem) may be employed in order to counteract the noise present in fMRI data, and which can facilitate the determination of when a signal was truly highly influenced by the task. This regularization is driven by the difference of the task frequency power to other power of the frequencies and minimizes the following:

|L4x - 2>||² + ||Γχ||²

with the solution of:

x = (A^TA + Γ^ΤΓ) ^~1A^Tb

where b is a vector of the ideal difference of power spectra values favouring the task frequency: the absolute difference of all power spectra values from that of the task frequency where the task frequency power is 1 . Since the power spectra is normalized, all other powers other than task would be 0. In other words, b is a vector of the absolute differences of power spectra values from that of the task frequency power where the task frequency is the only frequency present, i.e., if Pt would equal 1 and is the power of the task frequency, so if

abs(P₂-Pt) ... abs(Pt-Pt) ... abs(PN-Pt)], then J =[1 1 ... 0 ...1 ].

The matrix A is the identity matrix I multiplied by the actual differences of task frequency power to all other frequencies' power (e.g., An is the absolute difference of power of the first frequency and power of the task frequency). Lastly, Γ is the identity matrix multiplied by a constant a, where smaller values of a result in stronger regularization. The vector x is multiplied by the task frequency power difference values, and the resulting values of this multiplication are subtracted from the original task frequency power to create a new power spectrum. Next, the minimum value is subtracted and the new power spectrum is renormalized. Tikhonov thusly regularization provides data- driven control to the regularization.

It is noted that although Tikhonov regularization has previously been applied to fMRI , it has been done so in very different contexts, and to the best of the knowledge of the present inventors, has never been applied to improve calculation of fMRI data entropy, nor has ever been applied for the

determination of a metric of fMRI data quality. Previously, Tikhonov regularization has been used in fMRI for three primary applications. The first application is to improve reconstruction of under-sampled images in fMRI acquisitions (see, e.g., Ying, Lu & Liang, IEEE transactions on medical imaging 19.12 (2000): 1 188-1201 ). The second is improving the

extraction/classification of fMRI spatial activation patterns (see, e.g.,

Bharathan, Arun K. , et al. Green Computing, Communication and

Conservation of Energy (ICGCE), 2013 International Conference on. IEEE, 2013). The third is modelling of the hemodynamic response function in fMRI (see, e.g, Vakorin, Vasily A., Ron Borowsky, and Gordon E. Sarty. Statistics in medicine26.21 (2007): 3830-3844.). As explained above, use of Tikhonov regularization for improving entropy calculation of fMRI responses has not been previously reported.

The task frequency is essential to the regularization in this usage. The point of the regularization is to make data sparser in the frequency domain, but only in favour of the task frequency. Otherwise, a sparse power spectrum could be defined by another frequency, including that of obvious known noise sources or low frequency drift. Therefore the regularization is based around differences of powers in the frequency domain. The nature of Tikhonov regularization is such that it allows the regularization to be data driven, using the task frequency prior knowledge, such that it is sensitive to the relative power of the task frequency to that of other frequencies, as shown below.

Spectral entropy is calculated using the frequency powers from the fast Fourier transform (FFT):

where I = 1 : N, and N is the total number of power spectra frequencies. The i^th frequency power is given by F_t .

A demonstration of the effect of the regularization on spectral entropy is shown in FIGS. 2A-C. These plots demonstrate how the effect of regularization increases with the difference of the power of the task frequency to the power of all other frequencies. The original power spectra are shown in blue, and the regularized power spectra are shown in red. In FIG. 2A, the task frequency power is comparable to the other frequencies' powers and the regularization has little effect. In FIG. 2B and FIG. 2C, the task frequency dominates the power spectra and the regularization exaggerates this effect, with FIG. 2C exhibiting a larger power difference than FIG. 2B. The larger the difference between the power of the task frequency and other frequencies, the greater the regularization as evidenced by comparing FIG 2B and 2C. In all plots, the same regularization value, alpha, of 0.07 was used.

Avoiding under or over-regularization is important for the use of spectral entropy for the calculation of fMRI quality measures. According to one example embodiment, in order to determine the appropriate strength of regularization parameter a, simulations (fMRI training data) with known variations in the effect size of the synthetic functional activation (achieved by varying the percent signal change during activation, relative to the mean "resting" signal) may be employed such that the "ground truth" of activation could be known and used to inform the calculation. In the work shown here, 48 simulated data sets with varying percent signal change and re-sampled noise were generated for use in determining the regularization strength, as described below.

In one example implementation, in order to determine a suitable regularization strength value a, simulated data sets [Welvaert, Marijke, et al. "neuRosim: An R package for generating fMRI data." Journal of Statistical Software 44.10 (201 1 ): 1 -18] with percent signal change ranging from 2-5% in approximately the right motor cortex were created with a ranging from 0.005 to 0.15. First degree spline interpolation was performed between the a values to minimize the need for additional simulated data sets and give finer resolution of a. The choice of a suitable value for a was determined as the value that optimized sensitivity and specificity. In the present example implementation, this maximized sensitivity and specificity was determined based on distance to the top left corner of a receiver-operator curve (ROC) plot) across the range of percent signal change as well as the minimization of the standard deviation of the optimal sensitivity and specificity across all percent signal changes. In the present example implementation, the distance to top left corner of the ROC and standard deviation were equally weighted in a cost function that was employed to determine the value of the regularization parameter. This non-limiting approach was employed in the present example implementation because optimal sensitivity and specificity were considered comparably important to consistent performance of the regularization across different levels of activation.

FIG. 3A shows representative results of this data-driven optimization process for selecting regularization, as determined for a specific simulation data set. The regularization values may be determined based on a large number of simulated data sets with varying activation patterns, so as to ensure generalizability. FIG. 3B plots the error for the plot shown in FIG. 3A, calculated using 5 sets of simulations with the same levels of activation and the same ROI. Low error compared to actual values gives confidence.

The selected regularization may then be employed in the calculation of spectral entropy. As shown in the examples below, the regularization of the spectral entropy has been found to provide improved spectral entropy and fMRI data quality measures.

The regularized entropy described here, determined on a whole-brain voxel-wise basis, may be converted to quality metrics in a variety of ways. These include, but are not limited to the following.

In one example embodiment, by plotting a histogram of the voxel-wise entropy values, it is possible to characterize changes in the shape of the whole-brain distribution. Such changes in shape are correlated to changes in data quality. In one example implementation, a shift in the whole-brain distribution of regularized frequency-domain entropy to lower values can be employed as indicative of an increase in task information compared to noise (a measure of data quality). In other example implementations, one could calculate the spatial extent of clusters of low entropy voxels, given that real activation does not typically exist as individual voxels and therefore exists as clusters of low entropy voxels.

According to some example embodiments, any and/or all of these measures can then be compared to a template frequency-domain (e.g. spectral) entropy distribution, which can be age dependent, by averaging many healthy human brains at rest. This is similar to the way that the "MNI Brain", which averages images of 152 brains, is used to examine differences in brain anatomy compared to the template. Doing so permits calculation of the variances in these measures that describe differences when a brain is active versus at rest, which represents a metric for task information data quality. In other example implementations, machine learning can also be used to create classifiers for individual entropy distributions, so as to classify fMRI data sets with high amounts of task information compared to noise, which is indicative of overall data quality.

In another example implementation, a measure that can be used to create a data quality metric is the creation of an individual level "bootstrap" spectral entropy distribution. A bootstrapped fMRI data set can be created from a single time-course extracted from real fMRI data acquired while a subject is performing a task. The time-course from a voxel within a region of activation is then re-sampled repeatedly and randomly to create a new time- course for each voxel in a template brain. The voxel may be sampled with and without repeats (i.e. the same time point may or may not be included more than once). The resulting data set will be therefore made entirely of resampled bootstrap signals, and any task information that remains in the signals is purely due to random chance. This would allow creation of a "standard entropy distribution" with particular regularization parameters, one that would be expected from purely random signals that may or may not contain any trace of task and could therefore be compared to real distributions from data that we wish to characterize with an entropy metric, so as to quantify the amount of task information present.

FIGS. 4A and 4B plot the spectral entropy values with and without regularization (the value of which was chosen using the method described above), only for voxels located within the active region of the simulated fMRI data sets. This data was generated from 720 individual spectral entropy maps, generated with step-wise varying regularization applied to 48 different simulated fMRI data sets. Given that they are "active", the voxels should be low in entropy, with some variation due to differences in data quality due to varying percent signal change between 2-5%. However, without

regularization, the values are widely spread with most having an entropy > 0.5. With regularization, the spectral entropy of the active voxels is shown to greatly decrease, with many voxels showing very low entropy, as one would expect from an active region.

Furthermore, the regularization should also selectively lower the entropy of the task signal and not the noise signal over a reasonable range of regularization values. This is demonstrated in FIGS. 5A-C, which shows that the specific form of regularization primarily affects only the task signal, and only begins to significantly affect noise signals when over-regularized (i.e. for very low alpha). Indeed, FIGS. 5A-C demonstrate how spectral entropy changes much more dramatically with regularization strength for a signal containing task information than a signal that is purely noise. In FIG. 5A, the noisy signal's spectral entropy is relatively invariant, which is desired, but the task signal changes considerably. This demonstrates the data-driven nature of the regularization method. The task signal is shown in FIG. 5B, and the noisy signal is shown in FIG. 5C.

Although many of the example embodiments disclosed herein employ Tikhonov regularization for the regularization of the fMRI power spectrum, it will be understood that Tikhonov regularization is merely disclosed as an example form of regularization. The present inventors selected Tikhonov regularization due to its known effectiveness in regularizing ill-posed problems, but other forms of regularization may be employed in the alternative, which have relative strengths and weaknesses. These may include, but are not limited to forms of L1 and L2 regularization of the frequency domain data, or a combination of both. L1 regularization methods are based on the minimization of the sum of absolute differences between target and estimated values and are known to be robust, but may produce multiple regularization solutions, as the L1-norm does not necessarily have an analytical solution. The task frequency would again be employed as the element in a vector containing all of the frequency content information. When used with L1 regularization, it would be expressed as difference values of frequencies in the spectrum to that of the task, with the minimization occurring between the actual difference values and a weighted vector of ideal differences of a perfectly sparse power spectrum favoring the task frequency.

Furthermore, although many of the example embodiments described herein refer to the regularization of spectral entropy as calculated by Shannon entropy in the frequency domain, it will be understood that Shannon entropy in the frequency domain is merely disclosed as a non-limiting example of a wide variety of suitable entropy measures, and that other entropy measures may be used in alternative implementations of the examples disclosed herein. For example, although Shannon entropy is defined as employing a base 2 logarithm, spectral entropy could alternatively be calculated with a different log base.

It is also noted that other measures of entropy may be employed in the frequency domain, and that spectral entropy is described herein merely as a suitable example due to its simplicity of formulation and speed of calculation. For example, in one alternative implementation, the Gini index may be employed the frequency domain, and may be formulated differently (i.e. the usage of "low" and "high" entropy would need to be reversed such that it represented sparsity of information clustered in a single frequency coefficient in the power spectrum). Renyi entropy, an index for information "diversity" and the generalization of Hartley entropy, might also be used as it does not necessitate the use of probabilities and actually converges to Shannon's entropy. The probabilities of Renyi entropy would be replaced by frequency powers, not unlike the transformation of Shannon's entropy to spectral entropy. Renyi entropy would represent a descriptor of the amount of randomness (diversity) in the frequency distributions rather than being a measure of average information content. Accordingly, it will be understood that a variety of different frequency-domain measures of entropy may be employed to quantify frequency sparsity.

Referring now to FIG. 6, a flow-chart is provided that describes an example method of generating an fMRI data quality metric based on a calculation of regularized (e.g. Tikhonov regularized) frequency-domain entropy (e.g. spectral entropy). A magnetic resonance imaging scanner is employed to perform scanning of a subject according to a selected a functional magnetic resonance protocol (e.g. a block design paradigm), generating raw fMRI data as shown at 100. The raw fMRI data may be preprocessed as shown at steps 105, 1 10 and 1 15, including skull stripping, motion realignment, and low-pass temporal filtering, respectively. As shown at 120, a regularization parameter, selected according to the example methods described above (e.g. selection of a regularization parameter, for use with spectral-domain Tikhonov regularization, that maximizes the sensitivity and/or specificity based on one or more fMRI training or simulation datasets), is obtained. This regularization parameter is employed to perform regularization of the power spectrum, as shown at 125. The regularized power spectrum is then processed to calculate regularized frequency-domain entropy at step 130. The regularized entropy, determined on a per-voxel level, is employed for the generation of one or more fMRI data quality metrics. For example, as shown in FIG. 6, a regularized entropy histogram may be generated as shown at 135, and processed to generate one or more distribution measures characterizing the histogram, as shown at 135.

These measures may be employed as fMRI data quality metrics, and/or further processed to generate one or more fMRI data quality metrics, as shown at step 145, the one or more fMRI data quality metrics may be reported. A determination may then be made as to whether or not the fMRI data is of sufficient quality. For example, this determination may be

automated, according to pre-selected criterion associated with the one or more fMRI data quality metrics. Alternatively, this determination may be based on user intervention, e.g. based on a user or operator reviewing fMRI data quality metrics such as the spectral entropy histogram and/or measures associated with the calculated spectral entropy.

In the event that a determination is made that the fMRI data is of sufficient quality, the process may terminate, as shown at 155. On the other hand, if one or more fMRI data metrics fail to satisfy pre-selected criteria and/or user acceptance, additional scanning may be performed, and additional fMRI raw data may be collected at shown at 160. This additional data may be pre-processed according to one or more of steps 165 to 175. In one example embodiment, the additional fMRI data may be appended to previously collected fMRI data, as shown at step 180. Alternatively, the previously obtained fMRI data may be replaced with the additionally acquired fMRI data. Steps 125 to 150 may then be repeated, whereby newly determined fMRI data quality metrics are assessed at step 150. Additional data may subsequently be obtained one or more times until fMRI data of sufficient quality has been collected.

Referring now to FIG. 7, an example system is illustrated for performing functional magnetic resonance imaging according to the example methods described above. The example system includes a magnetic resonance scanner 50 that employs a main magnet 52 to produce a main magnetic field B0, which generates a polarization in a patient 60 or the examined subject. The example system includes gradient coils 54 for generating magnetic field gradients. A receive coil 58 detects RF signals from patient 60. The receive coil 58 can also be used as a transmission coil for the generation of radio frequency (RF) pulses. Alternatively, a body coil 56 may be employed to radiate and/or detect RF pulses. The RF pulses are generated by an RF unit 65, and the magnetic field gradients are generated by a gradient unit 70.

It will be understood that the MR system can have additional units or components that are not shown for clarity, such as, but not limited to, additional control or input devices, and additional sensing devices, such as devices for cardiac and/or respiratory gating. Furthermore, the various units can be realized other than in the depicted separation of the individual units. It is possible that the different components are assembled into units or that different units are combined with one another. Various units (depicted as functional units) can be designed as hardware, software or a combination of hardware and software.

In the example system shown in FIG. 7, a control and processing hardware 200 controls the MRI scanner to generate RF pulses according to a suitable pulse sequence. The control and processing hardware 200 is interfaced with the MRI scanner 50 for controlling the acquisition of the received MRI signals. The control and processing hardware 200 acquires the received MRI signals from the RF unit 65 and processes the MRI signals according to the methods described herein in order to perform functional magnetic resonance imaging, generate raw fMRI image data, and processing the raw fMRI data in order to generate and optionally assess one or more fMRI data quality metrics.

The control and processing hardware 200 may be programmed with a set of instructions which when executed in the processor causes the system to perform one or more methods described in the present disclosure. For example, as shown in FIG. 7, control and processing hardware 200 may be programmed with instructions in the form of a set of executable image processing modules, such as, but not limited to, a pulse sequence generation module 245, an image reconstruction module 250, and fMRI data quality metric module 255. The pulse sequence generation module 245 may be implemented using algorithms known to those skilled in the art for pulse sequence generation, such as those described above.

During MRI scanning, RF data is received from the RF coils 56 and/or 58. The pulse sequence generation module 245 establishes the sequence of RF pulses and magnetic field gradients depending on the desired imaging sequence, MR signals responsively emitted by the patient and detected by the coils 56 and/or 58 are acquired. The image reconstruction module 245 processes the acquired MRI signals to perform image reconstruction and MRI image generation. The fMRI data quality metric module generates and optionally assesses one or more fMRI data quality metrics according to the example method shown in FIG. 6, or variations thereof.

The control and processing hardware 200 may include, for example, one or more processors 210, memory 215, a system bus 205, one or more input/output devices 220, and a plurality of optional additional devices such as communications interface 235, data acquisition interface 240, display 225, and external storage 230.

It is to be understood that the example system shown in FIG. 7 is illustrative of a non-limiting example embodiment, and is not intended to be limited to the components shown. For example, the system may include one or more additional processors and memory devices. Furthermore, one or more components of control and processing hardware 200 may be provided as an external component that is interfaced to a processing device.

Some aspects of the present disclosure can be embodied, at least in part, in software, which, when executed on a computing system, configures the computing system as a specialty-purpose computing system that is capable of performing the signal processing and noise reduction methods disclosed herein, or variations thereof. That is, the techniques can be carried out in a computer system or other data processing system in response to its processor, such as a microprocessor, CPU or GPU, executing sequences of instructions contained in a memory, such as ROM, volatile RAM, non-volatile memory, cache, magnetic and optical disks, cloud processors, or other remote storage devices. Further, the instructions can be downloaded into a computing device over a data network, such as in a form of a compiled and linked version. Alternatively, the logic to perform the processes as discussed above could be implemented in additional computer and/or machine readable media, such as discrete hardware components as large-scale integrated circuits (LSI's), application-specific integrated circuits (ASIC's), or firmware such as electrically erasable programmable read-only memory (EEPROM's) and field-programmable gate arrays (FPGAs).

A computer readable medium can be used to store software and data which when executed by a data processing system causes the system to perform various methods. The executable software and data can be stored in various places including for example ROM, volatile RAM, non-volatile memory and/or cache. Portions of this software and/or data can be stored in any one of these storage devices. In general, a machine-readable medium includes any mechanism that provides (i.e., stores and/or transmits) information in a form accessible by a machine (e.g., a computer, network device, personal digital assistant, manufacturing tool, any device with a set of one or more processors, etc.).

Examples of computer-readable media include but are not limited to recordable and non-recordable type media such as volatile and non-volatile memory devices, read only memory (ROM), random access memory (RAM), flash memory devices, floppy and other removable disks, magnetic disk storage media, optical storage media (e.g., compact discs (CDs), digital versatile disks (DVDs), etc.), network attached storage, cloud storage, among others. The instructions can be embodied in digital and analog communication links for electrical, optical, acoustical or other forms of propagated signals, such as carrier waves, infrared signals, digital signals, and the like. As used herein, the phrases "computer readable material" and "computer readable storage medium" refer to all computer-readable media, except for a transitory propagating signal per se.

The Tikhonov regularized spectral entropy, which is an example of a regularized frequency-domain entropy measure, and which is a non- parametric and model-free measure, is shown in the following examples to be a suitable metric for data quality, as demonstrated through a variety of measures. The ability of regularized spectral entropy to be rapidly calculated as a non-parametric almost-model-free value, with minimal pre-processing of raw fMRI data, that demonstrates a sensitive linearity with differences in data quality, validates its use as quality assurance metric for fMRI that can be implemented either in real-time or immediately following the scan completion. EXAMPLES

The following examples are presented to enable those skilled in the art to understand and to practice embodiments of the present disclosure. They should not be considered as a limitation on the scope of the disclosure, but merely as being illustrative and representative thereof.

Example 1 : Correlation of Spectral Entropy with Functional Contrast

FIG. 8 is a plot of showing the relationship between spectral entropy and the percent signal change, which is a form of fMRI data quality. Increased percent signal change indicates a higher amount of task information relative to noise, and hence higher sparsity in the frequency domain. The figure demonstrates how regularized spectral entropy decreases with increasing percent signal change, exhibiting an approximately linear relationship between regularized spectral entropy and percent signal change. That is, as the functional contrast weakens from 5 to 2%, the regularized spectral entropy increases approximately linearly over this range. All data shown in this plot is averaged from within the simulated "active" region, as all voxels outside this region will have a percent signal change of zero.

Example 2: Correlation of Spectral Entropy with Parametric T-Test

Referring now to FIG. 9, this figure demonstrates how regularized spectral entropy decreases with increasing t-statistic in the active ROI of simulations of varying percent signal change. The plot demonstrates an approximately linear relationship between regularized spectral entropy and conventional t-statistic. The regularized spectral entropy is thus shown to be correlated with the conventional parametric t-test (i.e. the statistical correlation to the model time course). All data shown in this plot is averaged from within the simulated "active" region, as all voxels outside this region will have a percent signal change of zero, and hence a low t-stat. The data set used in FIG. 9 is the same as that used to generate FIG. 8, which was created using 48 different simulated data sets (with 8 varying percent signal changes, and each re-sampled 6 times to generate different noise profiles).

FIG. 1 1 plots the sensitivity and specificity of parametric maps (GLM) and regularized spectral entropy as a function of percent signal change. Error bars are from 5 replications of each simulated data set. Spectral entropy had sensitivity of 0.7296 +/- 0.0086, and specificity of 0.7532 +/- 0.005 for 32 simulations of varying percent signal change. All data determined using the simulated data sets described previously. While the GLM, as expected, has better sensitivity and specificity, it is also important to note is that the GLM performs better in terms of sensitivity and specificity, but it also requires more assumptions and a priori knowledge, computation time, and pre-processing.

FIG. 12 shows an example of an fMRI language scan from a representative subject. Regions in the lightest grey have low regularized spectral entropy (below 0.2), regions in medium grey have high t-statistic (above 7) found using the GLM, and regions in the darkest grey overlaid on the brain image are overlap of low regularized spectral entropy and high t- statistic. A high degree of overlap occurs in a language task activated region in the left cortex.

Referring now to FIGS. 13 and 13B, t-statistic histograms were generated using the real fMRI data, with voxels grouped based on regularized spectral entropy values. Real data used to inform noise parameters in the simulations was collected on a General Electric 3-Tesla MR750 Discovery MRI using an echo- planar imaging sequence. Acquisition time for the anatomical scans were approximately 5 minutes. Subject data was collected with informed consent and Nova Scotia Heath Authority Research Ethics Board approval, using an integrated fMRI language paradigm.

The language scan included three tasks: word generation to a presented letter, generating the missing word at the end of incomplete sentences, and naming items or concepts described by a sentence. Contrast included scanning two patterns on either side of a fixation cross to determine if the patterns were the same, and alternating finger tapping. The latter contrast is intended to mimic the visual effect of reading a sentence, while the former counteracts any unintended motion induced by thinking of speech or motion-related words.

There were a total of 9 task blocks of length 24 seconds and 8 contrast blocks with length of 24 seconds. Before each block, a 1 -second warning was displayed for both contrast and task to "prime" the subject of the incoming stimulus. The total paradigm length, including a warmup time of 6 TR's, was 7 minutes and 12 seconds.

High values of regularized spectral entropy are shown in FIG. 13A, and low regularized spectral entropy is shown in FIG. 13B. Because of the large number of voxels, most of which do not have low entropy or high t-stat, histograms of low and high regularized spectral entropy vs. t-stat were made instead. By dividing the data into low and high regularized spectral entropy, it is more presentable and easier to interpret.

Low values of regularized spectral entropy favors more extreme values of t-statistic, indicating a relationship to the amount of task-related information content. Voxels with a regularized spectral entropy value greater than 0.2 were shown to have a corresponding t-stat that tightly centered on zero, whereas those less than 0.2 favoured high t-stat values, which are more statistically likely to contain high quality task-active time-courses. This example therefore further supports the use of Tikhonov regularized spectral entropy for the generation of fMRI data quality metrics.

FIG. 14 shows the difference in regularized spectral entropy distributions on the same subject when performing the task (blue) and resting (red). The inset shows the greater fraction of low regularized spectral entropy voxels when the subject is actually performing the task.

Example 3: Replicability of Regularized Spectral Entropy as fMRI Data Quality Metric

FIG. 15 plots the mean value and deviation of average of the regularized spectral entropy distribution (whole brain) in four human subjects who successfully completed the same language task. The error between values is shown in light grey, and the average value in the dark line. The results shown a high degree of consistency, in particular in the lower entropy values, with only the higher entropy values representing noise exhibited higher intra-subject variability.

This example demonstrates consistency of regularized spectral entropy distributions and replicability between subjects, thereby demonstrating that regularized spectral entropy has a high degree of replicability as data metric. The specific embodiments described above have been shown by way of example, and it should be understood that these embodiments may be susceptible to various modifications and alternative forms. It should be further understood that the claims are not intended to be limited to the particular forms disclosed, but rather to cover all modifications, equivalents, and alternatives falling within the spirit and scope of this disclosure.

Claims

THEREFORE WHAT IS CLAIMED IS:

1 . A method of performing functional magnetic resonance imaging (fMRI) data quality assessment, the method comprising:

frequency-domain regularization of the power spectra of the patient fMRI dataset, wherein the regularization is dependent on the task frequency, thereby obtaining regularized power spectra;

f) generating frequency-domain entropy measures from the

2. The method according to claim 1 wherein, in the event that one or more fMRI data quality metrics do not meet pre-selected data quality criteria, repeating steps d) to g) one or more times until the pre-selected data quality criteria is met.

3. The method according to claim 1 or 2 wherein the regularization is performed according to Tikhonov regularization.

4. The method according to any one of claims 1 to 3 wherein the frequency- domain entropy is calculated as spectral entropy.

5. The method according to any one of claims 1 to 3 wherein the frequency- domain entropy is calculated as a frequency-domain measure of the Gini index.

6. The method according to any one of claims 1 to 5 wherein at least one fMRI data quality metric is based on a comparison with a template frequency- domain entropy distribution.

7. The method according to any one of claims 1 to 5 wherein at least one fMRI data quality metric is based on a shift a distribution of frequency-domain entropy to lower or higher values after regularization.

8. The method according to claim 7 wherein the quality metric is based on a fraction of voxels with regularized frequency-domain entropy above or below a cutoff.

9. The method according to any one of claims 1 to 5 wherein at least one fMRI data quality metric is based on a histogram of frequency-dependent entropy values.

10. The method according to claim 9 wherein the at least one fMRI data quality metric is based on one of a fraction of low frequency-dependent entropy values and one or more characteristics of a shape of the histogram.

1 1 . The method according to claim 10 wherein the one or more characteristics of the shape of the histogram include a height and/or position of one or more peaks.

12. The method according to any one of claims 1 to 5 wherein at least one fMRI data quality metric is based on statistical measure associated with a distribution of the frequency-domain entropy measures.

13. The method according to any one of claims 1 to 12 wherein the preselected performance criteria is based on measures of sensitivity and specificity.

14. The method according to any one of claims 1 to 5 wherein the preselected performance criteria is based on classifying measures derived from data collected with known levels of quality.

15. The system according to claim 14 wherein the data collected with known levels of quality is provided by introduction, in known quantity, of artifacts and/or data-quality-reducing measures.

16. A magnetic resonance imaging system for performing functional magnetic resonance imaging (fMRI), the system comprising:

a magnetic resonance imaging scanner; and

d) controlling the magnetic resonance imaging scanner to generate a sequence of RF pulses and detect RF signals that are responsively emitted by a subject during the performance of a series of tasks characterized by a task frequency, and processing the signals to generate a patient fMRI dataset; e) employing the selected regularization strength to perform

f) generating frequency-domain entropy measures from the regularized power spectra, thereby obtaining a set of regularized frequency- domain entropy measures; and

17. The system according to claim 16 wherein the control and processing hardware is configured such that, in the event that one or more fMRI data quality metrics do not meet pre-selected data quality criteria, repeating steps d) to g) one or more times until the pre-selected data quality criteria is met.

18. The system according to claim 16 or 17 wherein the regularization is performed according to Tikhonov regularization.

19. The system according to any one of claims 16 to 18 wherein the frequency-domain entropy is calculated as spectral entropy.

20. The system according to any one of claims 16 to 18 wherein the frequency-domain entropy is calculated as a frequency-domain measure of the Gini index.

21 . The system according to any one of claims 16 to 20 wherein at least one fMRI data quality metric is based on a comparison with a template frequency- domain entropy distribution.

22. The system according to any one of claims 16 to 20 wherein at least one fMRI data quality metric is based on a shift a distribution of frequency-domain entropy to lower values after regularization.

23. The system according to any one of claims 16 to 20 wherein at least one fMRI data quality metric is based on a histogram of frequency-dependent entropy values.

24. The system according to claim 23 wherein the at least one fMRI data quality metric is based on one of a fraction of low frequency-dependent entropy values and one or more characteristics of the of a shape of the histogram.

25. The system according to claim 24 wherein the one or more characteristics of the shape of the histogram include a height and/or position of one or more peaks.

26. The system according to any one of claims 16 to 20 wherein at least one fMRI data quality metric is based on statistical measure associated with a distribution of the frequency-domain entropy measures.

27. The system according to any one of claims 16 to 26 wherein the preselected performance criteria is based on measures of sensitivity and specificity.

28. The system according to any one of claims 16 to 26 wherein the preselected performance criteria is based on classifying measures derived from data collected with known levels of quality.

29. The system according to claim 28 wherein the data collected with known levels of quality is provided by introduction, in known quantity, of artifacts and/or data-quality-reducing measures.