Blog Archives

fMRI In Neuroscience: Efficiency of Event-related Experiment Designs

Event-related fMRI experiments are used to detect selectivity in the brain to stimuli presented over short durations. An event is generally modeled as an impulse function that occurs at the onset of the stimulus in question. Event-related designs are flexible in that many different classes of stimuli can be intermixed. These designs can minimize confounding behavioral effects due to subject adaptation or expectation. Furthermore, stimulus onsets can be modeled at frequencies that are shorter than the repetition time (TR) of the scanner. However, given such flexibility in design and modeling, how does one determine the schedule for presenting a series of stimuli? Do we space out stimulus onsets periodically across a scan period? Or do we randomize stimulus onsets? Furthermore what is the logic for or against either approach? Which approach is more efficient for gaining incite into the selectivity in the brain?

Simulating Two fMRI Experiments: Periodic and Random Stimulus Onsets

To get a better understanding of the problem of choosing efficient experiment design, let’s simulate two simple fMRI experiments. In the first experiment, a stimulus is presented periodically 20 times, once every 4 seconds, for a run of 80 seconds in duration. We then simulate a noiseless BOLD signal evoked in a voxel with a known HRF. In the second experiment, we simulate the noiseless BOLD signal evoked by 20 stimulus onsets that occur at random times over the course of the 80 second run duration.  The code for simulating the signals and displaying output are shown below:

rand('seed',12345);
randn('seed',12345);
TR = 1 % REPETITION TIME
t = 1:TR:20; % MEASUREMENTS
h = gampdf(t,6) + -.5*gampdf(t,10); % ACTUAL HRF
h = h/max(h); % SCALE TO MAX OF 1

% SOME CONSTANTS...
trPerStim = 4; % # TR PER STIMULUS FOR PERIODIC EXERIMENT
nRepeat = 20; % # OF TOTAL STIMULI SHOWN
nTRs = trPerStim*nRepeat
stimulusTrain0 = zeros(1,nTRs);

beta = 3; % SELECTIVITY/HRF GAIN

% SET UP TWO DIFFERENT STIMULUS PARADIGM...
% A. PERIODIC, NON-RANDOM STIMULUS ONSET TIMES
D_periodic = stimulusTrain0;
D_periodic(1:trPerStim:trPerStim*nRepeat) = 1;

% UNDERLYING MODEL FOR (A)
X_periodic = conv2(D_periodic,h);
X_periodic = X_periodic(1:nTRs);
y_periodic = X_periodic*beta;

% B. RANDOM, UNIFORMLY-DISTRIBUTED STIMULUS ONSET TIMES
D_random = stimulusTrain0;
randIdx = randperm(numel(stimulusTrain0)-5);
D_random(randIdx(1:nRepeat)) = 1;

% UNDERLYING MODEL FOR (B)
X_random = conv2(D_random,h);
X_random = X_random(1:nTRs);
y_random = X_random*beta;

% DISPLAY STIMULUS ONSETS AND EVOKED RESPONSES
% FOR EACH EXPERIMENT
figure
subplot(121)
stem(D_periodic,'k');
hold on;
plot(y_periodic,'r','linewidth',2);
xlabel('Time (TR)');
title(sprintf('Responses Evoked by\nPeriodic Stimulus Onset\nVariance=%1.2f',var(y_periodic)))

subplot(122)
stem(D_random,'k');
hold on;
plot(y_random,'r','linewidth',2);
xlabel('Time (TR)');
title(sprintf('Responses Evoked by\nRandom Stimulus Onset\nVariance=%1.2f',var(y_random)))
BOLD signals evoked by periodic (left) and random (right) stimulus onsets.

BOLD signals evoked by periodic (left) and random (right) stimulus onsets.

The black stick functions in the simulation output indicate the stimulus onsets and each red function is the simulated noiseless BOLD signal to those stimuli. The first thing to notice is the dramatically different variances of the BOLD signals evoked for the two stimulus presentation schedules. For the periodic stimuli, the BOLD signal quickly saturates, then oscillates around an effective baseline activation. The estimated variance of the periodic-based signal is 0.18. In contrast, the signal evoked by the random stimulus presentation schedule varies wildly, reaching a maximum amplitude that is roughly 2.5 times as large the maximum amplitude of the signal evoked by periodic stimuli. The estimated variance of the signal evoked by the random stimuli is 7.4, roughly 40 times the variance of the signal evoked by the periodic stimulus.

So which stimulus schedule allows us to better estimate the HRF and, more importantly, the amplitude of the HRF, as it is the amplitude that is the common proxy for voxel selectivity/activation? Below we repeat the above experiment 50 times. However, instead of simulating noiseless BOLD responses, we introduce 50 distinct, uncorrelated noise conditions, and from the simulated noisy responses, we estimate the HRF using an FIR basis set for each  repeated trial. We then compare the estimated HRFs across the 50 trials for the periodic and random stimulus presentation schedules. Note that for each trial, the noise is exactly the same for the two stimulus presentation schedules. Further, we simulate a selectivity/tuning gain of 3 times the maximum HRF amplitude and assume that the HRF to be estimated is 16 TRs/seconds in length. The simulation and output are below:

%% SIMULATE MULTIPLE TRIALS OF EACH EXPERIMENT
%% AND ESTIMATE THE HRF FOR EACH
%% (ASSUME THE VARIABLES DEFINED ABOVE ARE IN WORKSPACE)

% CREATE AN FIR DESIGN MATRIX
% FOR EACH EXPERIMENT
hrfLen = 16;  % WE ASSUME TO-BE-ESTIMATED HRF IS 16 TRS LONG

% CREATE FIR DESIGN MATRIX FOR THE PERIODIC STIMULI
X_FIR_periodic = zeros(nTRs,hrfLen);
onsets = find(D_periodic);
idxCols = 1:hrfLen;
for jO = 1:numel(onsets)
	idxRows = onsets(jO):onsets(jO)+hrfLen-1;
	for kR = 1:numel(idxRows);
		X_FIR_periodic(idxRows(kR),idxCols(kR)) = 1;
	end
end
X_FIR_periodic = X_FIR_periodic(1:nTRs,:);

% CREATE FIR DESIGN MATRIX FOR THE RANDOM STIMULI
X_FIR_random = zeros(nTRs,hrfLen);
onsets = find(D_random);
idxCols = 1:hrfLen;
for jO = 1:numel(onsets)
	idxRows = onsets(jO):onsets(jO)+hrfLen-1;
	for kR = 1:numel(idxRows);
		X_FIR_random(idxRows(kR),idxCols(kR)) = 1;
	end
end
X_FIR_random = X_FIR_random(1:nTRs,:);

% SIMULATE AND ESTIMATE HRF WEIGHTS VIA OLS
nTrials = 50;

% CREATE NOISE TO ADD TO SIGNALS
% NOTE: SAME NOISE CONDITIONS FOR BOTH EXPERIMENTS
noiseSTD = beta*2;
noise = bsxfun(@times,randn(nTrials,numel(X_periodic)),noiseSTD);

%% ESTIMATE HRF FROM PERIODIC STIMULUS TRIALS
beta_periodic = zeros(nTrials,hrfLen);
for iT = 1:nTrials
	y = y_periodic + noise(iT,:);
	beta_periodic(iT,:) = X_FIR_periodic\y';
end

% CALCULATE MEAN AND STANDARD ERROR OF HRF ESTIMATES
beta_periodic_mean = mean(beta_periodic);
beta_periodic_se = std(beta_periodic)/sqrt(nTrials);

%% ESTIMATE HRF FROM RANDOM STIMULUS TRIALS
beta_random = zeros(nTrials,hrfLen);
for iT = 1:nTrials
	y = y_random + noise(iT,:);
	beta_random(iT,:) = X_FIR_random\y';
end

% CALCULATE MEAN AND STANDARD ERROR OF HRF ESTIMATES
beta_random_mean = mean(beta_random);
beta_random_se = std(beta_random)/sqrt(nTrials);

% DISPLAY HRF ESTIMATES
figure
% ...FOR THE PERIODIC STIMULI
subplot(121);
hold on;
h0 = plot(h*beta,'k')
h1 = plot(beta_periodic_mean,'linewidth',2);
h2 = plot(beta_periodic_mean+beta_periodic_se,'r','linewidth',2);
plot(beta_periodic_mean-beta_periodic_se,'r','linewidth',2);
xlabel('Time (TR)')
legend([h0, h1,h2],'Actual HRF','Average \beta_{periodic}','Standard Error')
title('Periodic HRF Estimate')

% ...FOR THE RANDOMLY-PRESENTED STIMULI
subplot(122);
hold on;
h0 = plot(h*beta,'k');
h1 = plot(beta_random_mean,'linewidth',2);
h2 = plot(beta_random_mean+beta_random_se,'r','linewidth',2);
plot(beta_random_mean-beta_random_se,'r','linewidth',2);
xlabel('Time (TR)')
legend([h0,h1,h2],'Actual HRF','Average \beta_{random}','Standard Error')
title('Random HRF Estimate')
Estimated HRFs from 50 trials of periodic (left) and random (right) stimulus schedules

Estimated HRFs from 50 trials of periodic (left) and random (right) stimulus schedules

In the simulation outputs, the average HRF for the random stimulus presentation (right) closely follows the actual HRF tuning. Also, there is little variability of the HRF estimates, as is indicated by the small standard error estimates for each time points. As well, the selectivity/gain term is accurately recovered, giving a mean HRF with nearly the same amplitude as the underlying model. In contrast, the HRF estimated from the periodic-based experiment is much more variable, as indicated by the large standard error estimates. Such variability in the estimates of the HRF reduce our confidence in the estimate for any single trial. Additionally, the scale of the mean HRF estimate is off by nearly 30% of the actual value.

From these results, it is obvious that the random stimulus presentation rate gives rise to more accurate, and less variable estimates of the HRF function. What may not be so obvious is why this is the case, as there were the same number of stimuli and  the same number of signal measurements in each experiment. To get a better understanding of why this is occurring, let’s refer back to the variances of the evoked noiseless signals. These are the signals that are underlying the noisy signals used to estimate the HRF. When noise is added it impedes the detection of the underlying trends that are useful for estimating the HRF.  Thus it is important that the variance of the underlying signal is large compared to the noise so that the signal can be detected.

For the periodic stimulus presentation schedule, we saw that the variation in the BOLD signal was much smaller than the variation in the BOLD signals evoked during the randomly-presented stimuli. Thus the signal evoked by random stimulus schedule provide a better characterization of the underlying signal in the presence of the same amount of noise, and thus provide more information to estimate the HRF. With this in mind we can think of maximizing the efficiency of the an experiment design as maximizing the variance of the BOLD signals evoked by the experiment.

An Alternative Perspective: The Frequency Power Spectrum

Another helpful interpretation is based on a signal processing perspective. If we assume that neural activity is directly correspondent with the onset of a stimulus event, then we can interpret the train of stimulus onsets as a direct signal of the evoked neural activity. Furthermore, we can interpret the HRF as a low-pass-filter that acts to “smooth” the available neural signal in time. Each of these signals–the neural/stimulus signal and the HRF filtering signal–has with it an associated power spectrum. The power spectrum for a signal captures the amount of power per unit time that the signal has as a particular frequency \omega . The power spectrum for a discrete signal can be calculated from the discrete Fourier transform (DFT) of the signal F(\omega) as follows

P(\omega) = | F(\omega)|^2

Below, we use Matlab’s \text{fft.m} function to calculate the DFT and the associated power spectrum for each of the stimulus/neural signals, as well as the HRF.

%% POWER SPECTRUM ANALYSES
%% (ASSUME THE VARIABLES DEFINED ABOVE ARE IN WORKSPACE)

% MAKE SURE WE PAD SUFFICIENTLY
% FOR CIRCULAR CONVOLUTION
N = 2^nextpow2(nTRs + numel(h)-1);
nUnique = ceil(1+N/2); % TAKE ONLY POSITIVE SPECTRA

% CALCULATE POWER SPECTRUM FOR PERIODIC STIMULI EXPERIMENT
ft_D_periodic = fft(D_periodic,N)/N; % DFT
P_D_periodic = abs(ft_D_periodic).^2; % POWER
P_D_periodic = 2*P_D_periodic(2:nUnique-1); % REMOVE ZEROTH & NYQUIST

% CALCULATE POWER SPECTRUM FOR RANDOM STIMULI EXPERIMENT
ft_D_random = fft(D_random,N)/N; % DFT
P_D_random = abs(ft_D_random).^2; % POWER
P_D_random = 2*P_D_random(2:nUnique-1); % REMOVE ZEROTH & NYQUIST

% CALCULATE POWER SPECTRUM OF HRF
ft_h = fft(h,N)/N; % DFT
P_h = abs(ft_h).^2; % POWER
P_h = 2*P_h(2:nUnique-1); % REMOVE ZEROTH & NYQUIST

% CREATE A FREQUENCY SPACE FOR PLOTTING
F = 1/N*[1:N/2-1];

% DISPLAY STIMULI POWER SPECTRA
figure
subplot(131)
hhd = plot(F,P_D_periodic,'b','linewidth',2);
axis square; hold on;
hhr = plot(F,P_D_random,'g','linewidth',2);
xlim([0 .3]); xlabel('Frequency (Hz)');
set(gca,'Ytick',[]); ylabel('Magnitude');
legend([hhd,hhr],'Periodic','Random')
title('Stimulus Power, P_{stim}')

% DISPLAY HRF POWER SPECTRUM
subplot(132)
plot(F,P_h,'r','linewidth',2);
axis square
xlim([0 .3]); xlabel('Frequency (Hz)');
set(gca,'Ytick',[]); ylabel('Magnitude');
title('HRF Power, P_{HRF}')

% DISPLAY EVOKED SIGNAL POWER SPECTRA
subplot(133)
hhd = plot(F,P_D_periodic.*P_h,'b','linewidth',2);
hold on;
hhr = plot(F,P_D_random.*P_h,'g','linewidth',2);
axis square
xlim([0 .3]); xlabel('Frequency (Hz)');
set(gca,'Ytick',[]); ylabel('Magnitude');
legend([hhd,hhr],'Periodic','Random')
title('Signal Power, P_{stim}.*P_{HRF}')
Power spectrum of neural/stimulus (left), HRF (center), and evoked BOLD (right) signals

Power spectrum of neural/stimulus (left), HRF (center), and evoked BOLD (right) signals

On the left of the output we see the power spectra for the stimulus signals. The blue line corresponds to the spectrum for the periodic stimuli, and the green line the spectrum for the randomly-presented stimuli. The large peak in the blue spectrum corresponds to the majority of the stimulus power at 0.25 Hz for the periodic stimuli, as this the fundamental frequency of the periodic stimulus presentation (i.e. every 4 seconds). However, there is little power at any other stimulus frequencies. In contrast the green spectrum indicates that the random stimulus presentation has power at multiple frequencies.

If we interpret the HRF as a filter, then we can think of the HRF power spectrum as modulating the power spectrum of the neural signals to produce the power of the evoked BOLD signals. The power spectrum for the HRF is plotted in red in the center plot. Notice how a majority of the power for the HRF is at frequencies less than 0.1 Hz, and there is very little power at frequencies above 0.2 Hz. If the neural signal power is modulated by the HRF signal power, we see that there is little resultant power in the BOLD signals evoked by periodic stimulus presentation (blue spectrum in the right plot). In contrast, because the power for the neural signals evoked by random stimuli are spread across the frequency domain, there are a number of frequencies that overlap with those frequencies for which the HRF also has power. Thus after modulating neural/stimulus power with the HRF power, the spectrum of the BOLD signals evoked by the randomly-presented stimuli have much more power across the relevant frequency spectrum than those evoked by the periodic stimuli. This is indicated by the larger area under the green curve in the right plot.

Using the signal processing perspective allows us to directly gain perspective on the limitations of a particular experiment design which are rooted in the frequency spectrum of the HRF. Therefore, another way we can think of maximizing the efficiency of an experimental design is maximizing the amount of power in the resulting evoked BOLD responses.

Yet Another Perspective Based in Statistics: Efficiency Metric

Taking a statistics-based approach leads to a formal definition of efficiency, and further, a nice metric for testing the efficiency of an experimental design. Recall that when determining the shape of the HRF, a common approach is to use the GLM model

y = X \beta + \epsilon

Here y is the evoked BOLD signal and X is a design matrix that links a set of linear model parameters \beta to those responses. The variable \epsilon is a noise term that is unexplained by the model. Using an FIR basis formulation of the model, the weights in \beta represent the HRF to a stimulus condition.

Because fMRI data are a continuous time series, the underlying noise \epsilon is generally correlated in time. We can model this noise as a Gaussian process with zero mean and a constant multivariate covariance C_{\epsilon}. Note that this is analogous to the Generalized Least Squares (GLS) formulation of the GLM. In general, the values that comprise C_{\epsilon} are unknown and have to be estimated from the fMRI data themselves.

For a known or estimated noise covariance, the Maximum Likelihood Estimator (MLE) for the model parameters \beta(derivation not shown) is:

\hat \beta = (X^TC_{\epsilon}^{-1}X)X^TC_{\epsilon}^{-1}y

Because the ML estimator of the HRF is a linear combination of the design matrix X and a set of corresponding responses, which are both random variables (X can represent any possible experiment design, and y is by definition random), the estimator is itself a random variable. It thus follows that the estimate for the HRF also has a variance. (We demonstrated how \beta is a random variable in the 50 simulations above, where for each simulation X was held fixed, but due to the added noise y was a random variable. For each noise condition, the estimate for \beta took on different values.) We saw above how an HRF estimator with a large variance is undesirable, as it reduces our confidence in the estimates of the HRF shape and scale. Therefore we would like to determine an estimator that has a minimum overall variance.

A formal metric for efficiency of a least-squares estimator is directly related to the variance of the estimator. The efficiency is defined to be the inverse of the sum of the estimator variances. An estimator that has a large sum of variances will have a low efficiency, and vice versa. But how do we obtain the values of the variances for the estimator? The variances can be recovered from the diagonal elements of the estimator covariance matrix C_{\hat \beta}, giving the following definition for the efficiency, E

E = 1/trace(C_{\hat \beta})

In earlier post we found that the covariance matrix C_{\hat \beta} for the GLS estimator (i.e. the formulation above) with a given noise covariance C_{\epsilon} is:

C_{\hat \beta} = (X^T C_{\epsilon}^{-1} X)^{-1}.

Thus the efficiency for the HRF estimator is

E = 1/trace((X^T C_{\epsilon}^{-1}X)^{-1})

Here we see that the efficiency depends only on the known noise covariance (or an estimate of it), and the design matrix used in the model, but not the shape of the HRF. In general the noise covariance is out of the experimenter’s control (but see the take-homes below ), and must be dealt with post hoc. However, because the design matrix is directly related to the experimental design, the above expression gives a direct way to test the efficiency of experimental designs before they are ever used!

In the simulations above, the noise processes are drawn from an independent multivariate Gaussian distribution, therefore the noise covariance is equal to the identity (i.e. uncorrelated). We also estimated the HRF using the FIR basis set, thus our model design matrix was X_{FIR}. This gives the estimate the efficiency for the simulation experiments:

E_{simulation} = 1/trace(X_{FIR}^T X_{FIR})

Below we calculate the efficiency for the FIR estimates under the simulated experiments with periodic and random stimulus presentation designs.

%% ESTIMATE DESIGN EFFICIENCY
%% (ASSUME THE VARIABLES DEFINED ABOVE ARE IN WORKSPACE)

% CALCULATE EFFICIENCY OF PERIODIC EXPERIMENT
E_periodic = 1/trace(pinv(X_FIR_periodic'*X_FIR_periodic));

% CALCULATE EFFICIENCY OF RANDOM EXPERIMENT
E_random = 1/trace(pinv(X_FIR_random'*X_FIR_random));

% DISPLAY EFFICIENCY ESTIMATES
figure
bar([E_periodic,E_random]);
set(gca,'XTick',[1,2],'XTickLabel',{'E_periodic','E_random'});
title('Efficiency of Experimental Designs');
colormap hot;
Estimated efficiency for simulated periodic (left) and random (right) stimulus schedules.

Estimated efficiency for simulated periodic (left) and random (right) stimulus schedules.

Here we see that the efficiency metric does indeed indicate that the randomly-presented stimulus paradigm is far more efficient than the periodically-presented paradigm.

Wrapping Up

In this post we addressed the efficiency of an fMRI experiment design. A few take-homes from the discussion are:

  1. Randomize stimulus onset times. These onset times should take into account the low-pass characteristics (i.e. the power spectrum) of the HRF.
  2. Try to model selectivity to events that occur close in time. The reason for this is that noise covariances in fMRI are highly non-stationary. There are many sources of low-frequency physiological noise such as breathing, pulse, blood pressure, etc, all of which dramatically effect the noise in the fMRI timecourses. Thus any estimate of noise covariances from data recorded far apart in time will likely be erroneous.
  3. Check an experimental design against other candidate designs using the Efficiency metric.

Above there is mention of the effects of low-frequency physiological noise. Until now, our simulations have assumed that all noise is independent in time, greatly simplifying the picture of estimating HRFs and corresponding selectivity. However, in a later post we’ll address how to deal with more realistic time courses that are heavily influenced by sources of physiological noise. Additionally, we’ll tackle how to go about estimating the noise covariance C_{\epsilon} from more realistic fMRI time series.

Advertisements

fMRI in Neuroscience: Modeling the HRF With FIR Basis Functions

In the previous post on fMRI methods, we discussed how to model the selectivity of a voxel using the General Linear Model (GLM). One of the basic assumptions that we must make in order to use the GLM is that we also have an accurate model of the Hemodynamic Response Function (HRF) for the voxel. A common practice is to use a canonical HRF model established from previous empirical studies of fMRI timeseries. However, voxels throughout the brain and across subjects exhibit a variety of shapes, so the canonical model is often incorrect. Therefore it becomes necessary to estimate the shape of the HRF for each voxel.

There are a number of ways that have been developed for estimating HRFs, most of them are based on temporal basis function models. (For details on basis function models, see this previous post.). There are a number of basis function sets available, but in this post we’ll discuss modeling the HRF using a flexible basis set composed of a set of delayed impulses called Finite Impulse Response (FIR) basis.

Modeling HRFs With a Set of Time-delayed Impulses

Let’s say that we have an HRF with the following shape.

The HRF represented as the sum of the weighted FIR basis functions above.

A Model HRF.

We would like to be able to model the HRF as a weighted combination of simple basis functions. The simplest set of basis functions is the FIR basis, which is a series of H distinct unit-magnitude (i.e. equal to one) impulses, each of which is delayed in time by t = 1 \dots H TRs. An example of  modeling the HRF above using FIR basis functions is below:

%% REPRESENTING AN HRF WITH FIR BASIS FUNCTIONS
% CREATE ACTUAL HRF (AS MEASURED BY MRI SCANNER)
rand('seed',12345)
TR = 1                              % REPETITION TIME
t = 1:TR:20;                        % MEASUREMENTS
h = gampdf(t,6) + -.5*gampdf(t,10); % ACTUAL HRF
h = h/max(h);

% DISPLAY THE HRF
figure;
stem(t,h,'k','Linewidth',2)
axis square
xlabel(sprintf('Basis Function Contribution\nTo HRF'))
title(sprintf('HRF as a Series of \nWeighted FIR Basis Functions'))

% CREATE/DISPLAY FIR REPRESENTATION
figure; hold on
cnt = 1;

% COLORS BASIS FUNCTIONS ACCORDING TO HRF WEIGHT
map = jet(64);
cRange = linspace(min(h),max(h),64);

for iT = numel(h):-1:1
	firSignal = ones(size(h));
	firSignal(cnt) = 2;
	[~,cIdx] = min(abs(cRange-h(cnt)));
	color = map(cIdx,:);
	plot(1:numel(h),firSignal + 2*(iT-1),'Color',color,'Linewidth',2)
	cnt = cnt+1;
end
colormap(map); colorbar; caxis([min(h) max(h)]);

% DISPLAY
axis square;
ylabel('Basis Function')
xlabel('Time (TR)')
set(gca,'YTick',0:2:39,'YTickLabel',20:-1:1)
title(sprintf('Weighted FIR Basis\n Set (20 Functions)'));
Representing HRF as a weighted set of FIR basis functions. The color of each of the 20 basis functions corresponds to its weight

Representing the HRF above as a weighted set of FIR basis functions.
The color of each of the 20 basis functions corresponds to its weight

Each of the basis functions b_t has an unit impulse that occurs at time t = 1 \dots 20; otherwise it is equal to zero. Weighting each basis function b_t with the corresponding value of the HRF at each time point t, followed by a sum across all the functions gives the target HRF in the first plot above. The FIR basis model makes no assumptions about the shape of the  HRF–the weight applied to each basis function can take any value–which allows the model to capture a wide range of HRF profiles.

Given an experiment where various stimuli are presented to a subject and BOLD responses evoked within the subject’s brain, the goal is to determine the HRF to each of the stimuli within each voxel. Let’s take a look at a concrete example of how we can use the FIR basis to simultaneously estimate HRFs to many stimuli for multiple voxels with distint tuning properties.

Estimating the HRF of Simulated Voxels Using the FIR Basis

For this example we revisit a simulation of voxels with 4 different types of tuning (for details, see the previous post on fMRI in Neuroscience). One voxel is strongly tuned for visual stimuli (such as a light), the second voxel is weakly tuned for auditory stimuli (such as a tone), the third is moderately tuned for somatosensory stimuli (such as warmth applied to the palm), and the final voxel is unselective (i.e. weakly and equally selective for all three types of stimuli). We simulate an experiment where the blood-oxygen-level dependent (BOLD) signals evoked  in each voxel by a series of stimuli consisting of nonoverlapping lights, tones, and applications of warmth to the palm, are measured over T=330 fMRI measurments (TRs). Below is the simulation of the experiment and the resulting simulated BOLD signals:

%% SIMULATE AN EXPERIMENT
% SOME CONSTANTS
trPerStim = 30;
nRepeat = 10;
nTRs = trPerStim*nRepeat + length(h);
nCond = 3;
nVox = 4;
impulseTrain0 = zeros(1,nTRs);

% RANDOM ONSET TIMES (TRs)
onsetIdx = randperm(nTRs-length(h));

% VISUAL STIMULUS
impulseTrainLight = impulseTrain0;
impulseTrainLight(onsetIdx(1:nRepeat)) = 1;
onsetIdx(1:nRepeat) = [];

% AUDITORY STIMULUS
impulseTrainTone = impulseTrain0;
impulseTrainTone(onsetIdx(1:nRepeat)) = 1;
onsetIdx(1:nRepeat) = [];

% SOMATOSENSORY STIMULUS
impulseTrainHeat = impulseTrain0;
impulseTrainHeat(onsetIdx(1:nRepeat)) = 1;

% EXPERIMENT DESIGN / STIMULUS SEQUENCE
D = [impulseTrainLight',impulseTrainTone',impulseTrainHeat'];
X = conv2(D,h');
X = X(1:nTRs,:);

%% SIMULATE RESPONSES OF VOXELS WITH VARIOUS SELECTIVITIES
visualTuning =   [4 0 0]; % VISUAL VOXEL TUNING
auditoryTuning = [0 2 0]; % AUDITORY VOXEL TUNING
somatoTuning =   [0 0 3]; % SOMATOSENSORY VOXEL TUNING
noTuning =       [1 1 1]; % NON-SELECTIVE

beta = [visualTuning', ...
        auditoryTuning', ...
        somatoTuning', ...
        noTuning'];

y0 = X*beta;
SNR = 5;
noiseSTD = max(y0)/SNR;
noise = bsxfun(@times,randn(size(y0)),noiseSTD);
y = y0 + noise; % VOXEL RESPONSES

% DISPLAY VOXEL TIMECOURSES
voxNames = {'Visual','Auditory','Somat.','Unselective'};
cols = lines(4);
figure;
for iV = 1:4
	subplot(4,1,iV)
	plot(y(:,iV),'Color',cols(iV,:),'Linewidth',2); xlim([0,nTRs]);
	ylabel('BOLD Signal')
	legend(sprintf('%s Voxel',voxNames{iV}))
end
xlabel('Time (TR)')
set(gcf,'Position',[100,100,880,500])
Four simulated voxels, each with strong visual, weak auditory, moderate somatosensory and no tuning.

Four simulated voxels, each with strong visual, weak auditory,
moderate somatosensory and unselective tuning.

Now let’s estimate the HRF of each voxel to each of the C = 3 stimulus conditions using an FIR basis function model. To do so, we create a design matrix composed of successive sets of delayed impulses, where each set of impulses begins at the onset of each stimulus condition. For the [T \times C]-sized stimulus onset matrix D, we calculate an [T \times HC] FIR design matrix X_{FIR}, where H is the assumed length of the HRF we are trying to estimate. The code for creating and displaying the design matrix for an assumed HRF length H=16 is below:

%% ESTIMATE HRF USING FIR BASIS SET
% CREATE FIR DESIGN MATRIX
hrfLen = 16;  % WE ASSUME HRF IS 16 TRS LONG

% BASIS SET FOR EACH CONDITOIN IS A TRAIN OF INPULSES
X_FIR = zeros(nTRs,hrfLen*nCond);

for iC = 1:nCond
	onsets = find(D(:,iC));
	idxCols = (iC-1)*hrfLen+1:iC*hrfLen;
	for jO = 1:numel(onsets)
		idxRows = onsets(jO):onsets(jO)+hrfLen-1;
		for kR = 1:numel(idxRows);
			X_FIR(idxRows(kR),idxCols(kR)) = 1;
		end
	end
end

% DISPLAY
figure;
subplot(121);
imagesc(D);
colormap gray;
set(gca,'XTickLabel',{'Light','Tone','Som.'})
title('Stimulus Train');

subplot(122);
imagesc(X_FIR);
colormap gray;
title('FIR Design Matrix');
set(gca,'XTick',[8,24,40])
set(gca,'XTickLabel',{'Light','Tone','Som.'})
set(gcf,'Position',[100,100,550,400])

Left: The stimulus onset matrix (size = [T x 3]). Right the corresponding Design Matrix (size = [T x 3*H])

Left: The stimulus onset matrix (size = [T x 3]).
Right the corresponding Design Matrix (size = [T x 3*H])

In the right panel of the plot above, we see the  form of the FIR design matrix X_{FIR} for the stimulus onset on the left. For each voxel, we want to determine the weight on each column of X_{FIR} that will best explain the BOLD signals y measured from each voxel. We can form this problem in terms of a General Linear Model:

y = X_{FIR}\beta_{FIR}

Where \beta_{FIR} are the weights on each column of the FIR design matrix. If we set the values of \beta_{HRF} such as to minimize the sum of the squared errors (SSE) between the model above and the measured actual responses

SSE = \sum_i^N(y^{(i)} - X_{FIR}^{(i)})^2,

then we can use the Ordinary Least Squares (OLS) solution discussed earlier to solve the for \beta_{HRF}.  Specifically, we solve for the weights as:

\hat \beta_{FIR} = (X_{FIR}^T X_{FIR})^{-1} X_{FIR} y

Once determined, the resulting [CH \times V] matrix of weights \hat \beta_{FIR} has the HRF of each of the V=4 different voxels to each stimulus condition along its columns. The first H (1-16) of the weights along a column define the HRF to the first stimulus (the light). The second H (17-32) weights along a column determine the HRF to the second stimulus (the tone), etc… Below we parse out these weights and display the resulting HRFs for each voxel:

% ESTIMATE HRF FOR EACH CONDITION AND VOXEL
betaHatFIR = pinv(X_FIR'*X_FIR)*X_FIR'*y;

% RESHAPE HRFS
hHatFIR = reshape(betaHatFIR,hrfLen,nCond,nVox);

% DISPLAY
figure
cols = lines(4);
names = {'Visual','Auditory','Somat.','Unselective'};
for iV = 1:nVox
	subplot(2,2,iV)
	hold on;
	for jC = 1:nCond
		hl = plot(1:hrfLen,hHatFIR(:,jC,iV),'Linewidth',2);
		set(hl,'Color',cols(jC,:))
	end
	hl = plot(1:numel(h),h,'Linewidth',2);
	xlabel('TR')
	legend({'Light','Tone','Heat','True HRF'})
	set(hl,'Color','k')
	xlim([0 hrfLen])
	grid on
	axis tight
	title(sprintf('%s Voxel',names{iV}));
end
set(gcf,'Position',[100,100,880,500])
HRF estimates for each voxel to each of the 3 stimuli. Black plots show the shape of the actual HRF.

HRF estimates for each voxel to each of the 3 stimuli. Black plots show the shape of the true HRF.

Here we see that estimated HRFs accurately capture both the shape of the HRF and the selectivity of each of the voxels. For instance, the HRFs estimated from the responses of first voxel indicate strong tuning for the light stimulus. The HRF estimated for the light stimulus has an amplitude that is approximately 4 times that of the true HRF. This corresponds with the actual tuning of the voxel (compare this to the value of  \beta(1,1)). Additionally, time delay till the maximum value (time-to-peak) of the HRF to the light is the same as the true HRF. The first voxel’s HRFs estimated for the other stimuli are essentially noise around baseline. This (correctly) indicates that the first voxel has no selectivity for those stimuli. Further inspection of the remaining estimated HRFs indicate accurate tuning and HRF shape is recovered for the other three voxels as well.

Wrapping Up

In this post we discussed how to apply a simple basis function model (the FIR basis) to estimate the HRF profile and get an idea of the tuning of individual voxels. Though the FIR basis model can accurately model any HRF shape, it is often times too flexible. In scenarios where voxel signals are very noisy, the FIR basis model will tend to model the noise.

Additionally, the FIR basis set needs to incorporate a basis function for each time measurement.  For the example above, we assumed the HRF had a length of 16 TRs. The FIR basis therefore had 16 tuneable weights for each condition. This leads to a model with 48 (C\times H = 3 \times 16) tunable parameters for the GLM model. For experiments with many different stimulus conditions, the number of parameters can grow quickly (as HC). If the number of parameters is comparable (or more) than the number of BOLD signal measurements, it will be difficult accurately estimate \hat \beta_{FIR}. As we’ll see in later posts, we can often improve upon the FIR basis set by using more clever basis functions.

Another important but indirect issue that effects estimating the HRF is the experimental design, or rather the schedule used to present the stimuli. In the example above, the stimuli were presented in random, non-overlapping order. What if the stimuli were presented in the same order every time, with some set frequency? We’ll discuss in a later post the concept of design efficiency and how it affects our ability to characterize the shape of the HRF and, consequently, voxel selectivity.

fMRI in Neuroscience: Estimating Voxel Selectivity & the General Linear Model (GLM)

In a typical fMRI experiment  a series of stimuli are presented to an observer and evoked brain activity–in the form of blood-oxygen-level-dependent (BOLD) signals–are measured from tiny chunks of the brain called voxels. The task of the researcher is then to infer the tuning of the voxels to features in the presented stimuli based on the evoked BOLD signals. In order to make this inference quantitatively, it is necessary to have a  model of how BOLD signals are evoked in the presence of stimuli. In this post we’ll develop a model of evoked BOLD signals, and from this model recover the tuning of individual voxels measured during an fMRI experiment.

Modeling the Evoked BOLD Signals — The Stimulus and Design Matrices

Suppose we are running an event-related fMRI experiment where we present C different stimulus conditions to an observer while recording the BOLD signals evoked in their brain over a series of T consecutive fMRI measurements (TRs). We can represent the stimulus presentation quantitatively with a T \times C binary Stimulus Matrix, D, whose entries indicate the onset of each stimulus condition (columns) at each point in time (rows). Now let’s assume that we have an accurate model of how a voxel is activated by a single, very short stimulus. This activation model is called hemodynamic response function (HRF), h, for the voxel, and, as we’ll discuss in a later post, can be estimated from the measured BOLD signals. Let’s assume for now that the voxel is also activated to an equal degree to all stimuli. In this scenario we can represent the BOLD signal evoked over the entire experiment with another T \times C matrix X called the Design Matrix that is the convolution of the stimulus matrix D with the voxel’s HRF h.

X = D * h

Note that this model of the BOLD signal is an example of the Finite Impulse Response (FIR) model that was introduced in the previous post on fMRI Basics.

To make the concepts of D and X more concrete, let’s say our experiment consists of C = 3 different stimulus conditions: a light, a tone, and heat applied to the palm. Each stimulus condition is presented twice in a staggered manner during 80 TRs of fMRI measurements. The stimulus matrix and the design matrix are simulated here in Matlab:

TR = 1; % REPETITION TIME
t = 1:TR:20; % MEASUREMENTS
h = gampdf(t,6) + -.5*gampdf(t,10); % HRF MODEL
h = h/max(h); % SCALE HRF TO HAVE MAX AMPLITUDE OF 1

trPerStim = 30; % # TR PER STIMULUS
nRepeat = 2; % # OF STIMULUS REPEATES
nTRs = trPerStim*nRepeat + length(h);
impulseTrain0 = zeros(1,nTRs);

% VISUAL STIMULUS
impulseTrainLight = impulseTrain0;
impulseTrainLight(1:trPerStim:trPerStim*nRepeat) = 1;

% AUDITORY STIMULUS
impulseTrainTone = impulseTrain0;
impulseTrainTone(5:trPerStim:trPerStim*nRepeat) = 1;

% SOMATOSENSORY STIMULUS
impulseTrainHeat = impulseTrain0;
impulseTrainHeat(9:trPerStim:trPerStim*nRepeat) = 1;

% COMBINATION OF ALL STIMULI
impulseTrainAll = impulseTrainLight + impulseTrainTone + impulseTrainHeat;

% SIMULATE VOXELS WITH VARIOUS SELECTIVITIES
visualTuning = [4 0 0]; % VISUAL VOXEL TUNING
auditoryTuning = [0 2 0]; % AUDITORY VOXEL TUNING
somatoTuning = [0 0 3]; % SOMATOSENSORY VOXEL TUNING
noTuning = [1 1 1]; % NON-SELECTIVE

beta = [visualTuning', ...
        auditoryTuning', ...
        somatoTuning', ...
        noTuning'];

% EXPERIMENT DESIGN / STIMULUS SEQUENCE
D = [impulseTrainLight',impulseTrainTone',impulseTrainHeat'];

% CREATE DESIGN MATRIX FOR THE THREE STIMULI
X = conv2(D,h'); % X = D * h
X(nTRs+1:end,:) = []; % REMOVE EXCESS FROM CONVOLUTION

% DISPLAY STIMULUS AND DESIGN MATRICES
subplot(121); imagesc(D); colormap gray;
xlabel('Stimulus Condition')
ylabel('Time (TRs)');
title('Stimulus Train, D');
set(gca,'XTick',1:3); set(gca,'XTickLabel',{'Light','Tone','Heat'});

subplot(122);
imagesc(X);
xlabel('Stimulus Condition')
ylabel('Time (TRs)');
title('Design Matrix, X = D * h')
set(gca,'XTick',1:3); set(gca,'XTickLabel',{'Light','Tone','Heat'});

Stimulus presentation matrix, D (left) and the Design Matrix X for an experiment with three stimulus conditions: a light, a tone, and heat applied to the palm

Each column of the design matrix above (the right subpanel in the above figure) is essentially a model of the BOLD signal evoked independently by each stimulus condition, and the total signal is simply a sum of these independent signals.

Modeling Voxel Tuning — The Selectivity Matrix

In order to develop the concept of the design matrix we assumed that our theoretical voxel is equally tuned to all stimuli. However, few voxels in the brain exhibit such non-selective tuning. For instance, a voxel located in visual cortex will be more selective for the light than for the tone or the heat stimulus. A voxel in auditory cortex will be more selective for the tone than for the other two stimuli. A voxel in the somoatorsensory cortex will likely be more selective for the heat than the visual or auditory stimuli. How can we represent the tuning of these different voxels?

A simple way to model tuning to the stimulus conditions in an experiment is to multiplying each column of the design matrix by a weight that modulates the BOLD signal according to the presence of the corresponding stimulus condition. For example, we could model a visual cortex voxel by weighting the first column of X with a positive value, and the remaining two columns with much smaller values (or even negative values to model suppression). It turns out that we can model the selectivity of V individual voxels simultaneously through a C \times V Selectivity Matrix\beta. Each entry in \beta is the amount that the v-th voxel (columns) is tuned to the c-th stimulus condition (rows). Given the design matrix and the selectivity matrix, we can then predict the BOLD signals y of selectively-tuned voxels with a simple matrix multiplication:

y = X\beta

Keeping with our example experiment, let’s assume that we are modeling the selectivity of four different voxels: a strongly-tuned visual voxel, a moderately-tuned somatosensory voxel, a weakly tuned auditory voxel, and an unselective voxel that is very weakly tuned to all three stimulus conditions. We can represent the tuning of these four voxels with a 3 \times 4 selectivity matrix. Below we define a selectivity matrix that represents the tuning of these 4 theoretical voxels and simulate the evoked BOLD signals to our 3-stimulus experiment.

% SIMULATE NOISELESS VOXELS' BOLD SIGNAL
% (ASSUMING VARIABLES FROM ABOVE STILL IN WORKSPACE)
y0 = X*beta;

figure;
subplot(211);
imagesc(beta); colormap hot;
axis tight
ylabel('Condition')
set(gca,'YTickLabel',{'Visual','Auditory','Somato.'})
xlabel('Voxel');
set(gca,'XTick',1:4)
title('Voxel Selectivity, \beta')

subplot(212);
plot(y0,'Linewidth',2);
legend({'Visual Voxel','Auditory Voxel','Somato. Voxel','Unselective'});
xlabel('Time (TRs)'); ylabel('BOLD Signal');
title('Activity for Voxels with Different Stimulus Tuning')
set(gcf,'Position',[100 100 750 540])
subplot(211); colorbar

Selectivity matrix (top) for four theoretical voxels and GLM BOLD signals (bottom) for a simple experiment

The top subpanel in the simulation output visualizes the selectivity matrix defined for the four theoretical voxels. The bottom subpanel plots the columns of the T \times V matrix of voxel responses y. We see that the maximum response of the strongly-tuned visual voxel (plotted in blue) is larger than that of the other voxels, corresponding to the larger weight upper left of the selectivity matrix. Also note that the response for the unselective voxel (plotted in cyan) demonstrates the linearity property of the FIR model. The attenuated but complex BOLD signal from the unselective voxel results from the sum of small independent signals evoked by each stimulus.

Modeling Voxel Noise

The example above demonstrates how we can model BOLD signals evoked in noisless theoretical voxels. Though this noisless scenario is helpful for developing a modeling framework, real-world voxels exhibit variable amounts of noise (noise is any signal that cannot be accounted by the FIR model). Therefore we need to incorporate a noise term into our BOLD signal model.

The noise in a voxel is often modeled as a random variable \epsilon. A common choice for the noise model is a zero-mean Normal/Gaussian distribution with some variance \sigma^2:

\epsilon \sim \mathcal N(0,\sigma^2)

Though the variance of the noise model may not be known apriori, there are methods for estimating it from data. We’ll get to estimating noise variance in a later post when we discuss various sources of noise and how to account for them using more advance techniques. For simplicity, let’s just assume that the noise variance is 1 as we proceed.

Putting It All Together — The General Linear Model (GLM)

So far we have introduced on the concepts of the stimulus matrix, the HRF, the design matrix, selectivity matrix, and the noise model. We can combine all of these to compose a comprehensive quantitative model of BOLD signals measured from a set of voxels during an experiment:

y = X\beta + \epsilon \\ = (D * h)\beta + \epsilon

This is referred to as the General Linear Model (GLM).

In a typical fMRI experiment the researcher controls the stimulus presentation D, and measures the evoked BOLD responses y from a set of voxels. The problem then is to estimate the selectivities of the voxels based on these measurments. Specifically, we want to determine the parameters \hat \beta that best explain the measured BOLD signals during our experiment. The most common way to do this is a method known as Ordinary Least Squares (OLS) Regression. Using OLS the idea is to adjust the values of \hat \beta such that the predicted model BOLD signals are as similar to the measured signals as possible. In other words, the goal is to infer the selectivity each voxel would have to exhibit in order to produce the measured BOLD signals. I showed in an earlier post that the optimal OLS solution for the selectivities \hat \beta is given by:

\hat \beta = (X^T X)^{-1} X^T y

Therefore, given a design matrix X and a set of voxel responses y associated with the design matrix, we can calculate the selectivities of voxels to the stimulus conditions represented by the columns of the design matrix. This works even when the BOLD signals are noisy. To get a better idea of this process at work let’s look at a quick example based on our toy fMRI experiment.

Example: Recovering Voxel Selectivity Using OLS

Here the goal is to recover the selectivities of the four voxels in our toy experiment they have been corrupted with noise. First, we add noise to the voxel responses. In this example the variance of the added noise is based on a concept known as signal-to-noise-ration or SNR.  As the name suggests, SNR is the ratio of the underlying signal to the noise “on top of” the signal. SNR is a very important concept when interpreting fMRI analyses. If a voxel exhibits a low SNR, it will be far more difficult to estimate its tuning. Though there are many ways to define SNR, in this example it is defined as the ratio of the maximum signal amplitude to the variance of the noise model. The underlying noise model variance is adjusted to be one-fifth of the maximum amplitude of the BOLD signal, i.e. an SNR of 5.  Feel free to try different values of SNR by changing the value of the variable \text{SNR} in the Matlab simulation. Noisy versions of the 4 model BOLD signals are plotted in the top subpanel of the figure below. We see that the noisy signals are very different from the actual underlying BOLD signals.

Noisy BOLD signals from 4 voxels (top) and GLM predictions (bottom) of the underlying BOLD signals

Here we estimate the selectivities \hat \beta from the GLM using OLS, and then predict the BOLD signals in our experiment with this estimate. We see in the bottom subpanel of the above figure that the resulting GLM predictions of are quite accurate. We also compare the estimated selectivity matrix \hat \beta to the actual selectivity matrix \beta below. We see that OLS is able to recover the selectivity of all the voxels.

Actual (top) and estimated (bottom) selectivity matrices.

% SIMULATE NOISY VOXELS & ESTIMATE TUNING
% (ASSUMING VARIABLES FROM ABOVE STILL IN WORKSPACE)

SNR = 5; % (APPROX.) SIGNAL-TO-NOISE RATIO
noiseSTD = max(y0(:))./SNR; % NOISE LEVEL FOR EACH VOXEL
noise = bsxfun(@times,randn(size(y0)),noiseSTD);
y = y0 + noise;

betaHat = inv(X'*X)*X'*y % OLS
yHat = X*betaHat; % GLM PREDICTION

figure
subplot(211);
plot(y,'Linewidth',3);
xlabel('Time (s)'); ylabel('BOLD Signal');
legend({'Visual Voxel','Auditory Voxel','Somato. Voxel','Unselective'});
title('Noisy Voxel Responses');

subplot(212)
h1 = plot(y0,'Linewidth',3); hold on
h2 = plot(yHat,'-o');
legend([h1(end),h2(end)],{'Actual Responses','Predicted Responses'})
xlabel('Time (s)'); ylabel('BOLD Signal');
title('Model Predictions')
set(gcf,'Position',[100 100 750 540])

figure
subplot(211);
imagesc(beta); colormap hot(5);
axis tight
ylabel('Condition')
set(gca,'YTickLabel',{'Visual','Auditory','Somato.'})
xlabel('Voxel');
set(gca,'XTick',1:4)
title('Actual Selectivity, \beta')

subplot(212)
imagesc(betaHat); colormap hot(5);
axis tight
ylabel('Condition')
set(gca,'YTickLabel',{'Visual','Auditory','Somato.'})
xlabel('Voxel');
set(gca,'XTick',1:4)
title('Noisy Estimated Selectivity')
drawnow

Wrapping Up

Here we introduced the GLM commonly used for fMRI data analyses and used the GLM framework to recover the selectivities of simulated voxels. We saw that the GLM is quite powerful of recovering the selectivity in the presence of noise. However, there are a few details left out of the story.

First, we assumed that we had an accurate (albeit exact) model for each voxel’s HRF. This is generally not the case. In real-world scenarios the HRF is either assumed to have some canonical shape, or the shape of the HRF is estimated the experiment data. Though assuming a canonical HRF shape has been validated for block design studies of peripheral sensory areas, this assumption becomes dangerous when using event-related designs, or when studying other areas of the brain.

Additionally, we did not include any  physiological noise signals in our theoretical voxels. In real voxels, the BOLD signal changes due to physiological processes such as breathing and heartbeat can be far larger than the signal change due to underlying neural activation. It then becomes necessary to either account for the nuisance signals in the GLM framework, or remove them before using the model described above. In two upcoming posts we’ll discuss these two issues: estimating the HRF shape from data, and dealing with nuisance signals.