Investigation of human visual cortex responses to flickering light using functional near infrared spectroscopy and constrained ICA

The human visual sensitivity to the °ickering light has been under investigation for decades. The ̄nding of research in this area can contribute to the understanding of human visual system mechanism and visual disorders, and establishing diagnosis and treatment of diseases. The aim of this study is to investigate the e®ects of the °ickering light to the visual cortex by monitoring the hemodynamic responses of the brain with the functional near infrared spectroscopy (fNIRS) method. Since the acquired fNIRS signals are a®ected by physiological factors and measurement artifacts, constrained independent component analysis (cICA) was applied to extract the actual fNIRS responses from the obtained data. The experimental results revealed signi ̄cant changes ðp < 0:0001Þ of the hemodynamic responses of the visual cortex from the baseline when the °ickering stimulation was activated. With the uses of cICA, the contrast to noise ratio (CNR), re°ecting the contrast of hemodynamic concentration between rest and task, became larger. This indicated the improvement of the fNIRS signals when the noise was eliminated. In subsequent studies, statistical analysis was used to infer the correlation between the fNIRS signals and the visual stimulus. We found that there was a slight decrease of the oxygenated hemoglobin concentration (about 5:69%) over four frequencies when the modulation increased. However, the variations of oxy and deoxy-hemoglobin were not statistically signi ̄cant.


Introduction
Understanding the functioning of a visual system, from the eyes through visual cortex has progressed over centuries. Investigations to determine appropriate visual stimulations and data collection methods, and most importantly to understand the obtained responses have been widely reported in physiological and clinical research literature. [1][2][3] Photic stimulation using°ickering light that is modulated in sinusoidal waveform received particular attention. This is due to the fact that the stimulus parameters including modulation depth,°i cker frequency and average illuminance can be adjusted at will and independently. The responses are therefore rich in information and depend only on one parameter at a time. Furthermore, the experiments can be performed in vivo and noninvasively. Therefore, it has been broadly applied in the studies of visual system with animal as well as human. In psychophysics, sinusoidal stimulation has been used to establish human temporal modulation transfer functions (TMTF) 4 which is the sensitivity to°i ckering light of various frequencies. It has a typical band-pass shape and the critical°icker frequency (CFF) which is the highest°icker frequency one can perceive is about 50 Hz. They suggested that, the blood°ow in the optic nerve head is tightly coupled to neural activity. In electrophysiology, sinusoidal stimulation has been used to investigate electroretinograms (ERG), visual evoked potentials (VEP) and electrical responses of cells including photoreceptors, horizontal and ganglion cells. 5,6 These results revealed neural activities of di®erent visual cells at di®erent stages of the visual system. Toi and Riva 7 while investigating the retinal blood°ow at the optic nerve head of cats using laser Doppler method, found that sinusoidal stimulation always caused an increase in the blood°o w and the increase depended on the stimulus frequency, modulation depth and illuminance. When the modulation depth increased, the blood°o w increased and, most of the time, rapidly reached a saturation level as a sigmoidal function. The TMTF was established using a minimum blood°o w change as a criterion that had a band-pass shape and its high frequency slope was comparable to the slope obtained in the ganglion cell studies. 8 In functional magnetic resonance imaging (fMRI), an e®ective method used to measure the blood oxygen level detection (BOLD) signals, several studies revealed that the BOLD activation in the visual cortex increased when the stimulus frequency was raised up to 8 Hz. 9 Others found that BOLD activation peaked at 6-11 Hz. 10 However, fMRI is limited by its high cost and provides BOLD signals containing only deoxygenated hemoglobin information. Recently, functional near infrared spectroscopy (fNIRS) has emerged as an important modality applied to monitor the changes of oxygenated and deoxygenated hemoglobin (oxy-Hb and deoxy-Hb) concentrations. Due to its noninvasive and in vivo characteristics, fNIRS o®ers many advantages and becomes a complementary technique to fMRI. Typically, in fNIRS technology, a light within the near infrared spectrum from 700 to 900 nm wavelength is used to propagate through the brain matter. The wavelength is selected to maximize the absorption of the oxygenated and deoxygenated hemoglobin and minimize the light absorption of water and tissues. Therefore, fNIRS is able to monitor hemodynamic responses of the brain in real time.
It has been reported that fNIRS successfully monitored the hemodynamic responses at the visual cortex caused by°ickering stimulation that appears on a computer screen. 11 Meek et al. 12 suggested that oxy-Hb concentration increased whereas deoxy-Hb concentration decreased during stimulation. Kojima and Suzuki 13 came up with similar conclusion showing typical activation patterns of hemodynamic responses during a visual task, but suggested that intense focus to stimulus led to stronger changes on oxy-Hb concentration than passive watching. The location to measure hemodynamic responses of the visual cortex was exploited by Wijeakumar et al. 14 Liao et al. 15 revealed that neurovascular coupling in and around the visual cortex appeared not only in healthy adults but also in infants during the¯rst week of life. Bridge 16 used black and white concentric circles°ickering in bulls eye with spatial frequencies grating at di®erent temporal frequencies on the visual cortex. Experimental results showed insigni¯cant di®erences of the brain responses to various temporal frequencies of stimuli.
To the best of our knowledge, there has been no investigation using fNIRS to measure the hemodynamics of the brain when the eye is stimulated by sinusoidal waveform°ickering light of various frequencies and modulations depths. When the eyes are excited by sinusoidal°ickering light, neural signals are sent through optic nerve to stimulate the visual cortex. As a result, hemodynamic responses in this area may be changed and monitored by fNIRS to understand physiological and pathological principle of the ophthalmic system. This gives us useful insights on the e®ects of°ickering light on the operations of the visual cortex and may gain ben-e¯ts for the treatments of visual related diseases and disorders.
In this work, we investigated the variation of the oxy-Hb and deoxy-Hb changes at the visual cortex using fNIRS while the subjects observed a°ickering light modulated sinusoidally through an eye piece under Maxwellian view. 17 Since sensitivity of NIRS to brain activities is a®ected by artifacts originated from cardiac activities, Mayer waves and other physiology factors, methods to extract the actual NIRS responses from the measurement are important. Independent component analysis (ICA) constitutes a reliable method to recover source signals from mixtures and its procedure may extract task-related, physiological-related or artifact-related components. Incorporating prior constraints to ICA thereby plays an important role to isolate components targeted to speci¯c events in event-related experiments. 18 In this work, we proposed a method using constrained ICA (cICA) to extract the NIRS signals of interest from the mixture using reference signals. After the actual NIRS signals were recovered, we varied the frequencies and modulation depth of activated light to observe the corresponding NIRS responses.

Participants
A total of 10 young and healthy subjects (age average 20.6 years old and standard deviation 0.71) of both genders (four females and six males) participated in the experiments. None of them have the neurological disorders or the visual abnormalities. The local institutional review board has approved the studies and the participants have given written informed consent for the experimental contents. The tenets of the Declaration of Helsinki were followed.

Visual stimulator
The visual stimulator Papillometre [ Fig. 1(a)] has been described in details elsewhere. 17,[19][20][21] It generates a modulated light beam. The sinusoidal waveform of the beam is expressed as follows: where LðtÞ is the instantaneous light intensity, L 0 the average light illuminance ranging from 0 to 100 Trolands, m the modulation depth ranging from 0 to 1 and f the°ickering frequency ranging from 0 to 100 Hz. These parameters can be adjusted independently. The subject observed the stimuli under a Maxwellian view ranging from 5 to 60 . The stimulus color was white and¯eld of view was uniform.

fNIRS devices and data acquisition protocol
Our fNIRS equipment was the FOIRE-3000 (Shimadzu, Japan) which consisted of 16 optodes (8 light sources and 8 photo detectors) and constituted a maximum of 24 channels [ Fig. 1(b)]. The laser light sources have three di®erent wavelengths 780, 805 and 830 nm.
In our experiments performed here, the transmitter and receiver optodes are arranged in a 2 Â 3 matrix to cover the visual cortex at the back of the head [ Fig. 1(c)]. The T-transmitter are symbolized by a red circle whereas the R-receiver optodes by a blue one. Each pair of T-R with a distance of 3 cm constituted a measurement channel. Together, they formed seven channels.
The subjects sat in a quiet and dark room to avoid environmental distractions [ Fig. 1(d)]. Before performing the tasks, the subjects were¯rst asked to take a rest for 40 s to release the stress and to shift their hemodynamic signals into the baseline. Then, during the next 40 s, the subject performed the task by looking through the eyepiece of the Papillometre, a visual stimulator, to observe the stimuli. The fNIRS optodes of the FOIRE 3000, fNIRS equipment, were installed on a headgear which was secured on the subject's head at the visual cortex to monitor the changes of the oxy-Hb and deoxy-Hb concentrations at this area. The taskrest experimental sequence was repeated for¯ve times. Then, the subject will wait for several hours before he or she joins a new experimental session.
According to reported studies, 7,10 visual stimuli with frequencies lower than 15 Hz evoked strongest visual cortex responses. Therefore, we performed the experiments using four°ickering frequencies including 0, 5, 10 and 15 Hz with two modulation depths of 0.5 and 1. For each frequency and modulation depth, the experiment is repeated for two sessions. In total, each subject performed 5ðepochsÞ Â 2ðsessionsÞ Â 2ðmodulationsÞÂ 4ðfrequenciesÞ ¼ 80 NIRS trials.

Data processing
Since the optodes were arranged on the subjects' scalp and the limited distance that the light can travel deep into the brain matter was about 3 cm, 22 the obtained signals were the combination of the fNIRS signals and other artifacts such as Mayer waves, heartbeat and respiration movements and physiological signals appeared at the skin of the head.
To remove a part of noises, obtained NIRS signals are¯ltered by low-pass and high-pass¯lters with the cuto® frequency of 0.5 and 0.01 Hz. Previously correlation based signal improvement (CBSI) method 23 enhanced hemodynamic responses by an assumption of strong negative correlation between oxy-Hb and deoxy-Hb signals. Typically, when oxy-Hb concentration increases, the trends of deoxy-Hb concentration goes down accordingly. Let x 0 and y 0 be the measured fNIRS signals, then true fNIRS signals are recovered by CBSI using the following equation.

1450031-4
where ¼ stdðx 0 Þ stdðy 0 Þ is the ratio between the standard deviation of x 0 and of y 0 . Because the NIRS measurement is performed with the probes secured at the head skin, obtained signals are the mixtures of di®erent source signals emitted from active regions of the brain. Therefore, CBSI may magnify noninterest signals not involved to event-related paradigm in experiments.
Currently, it has been accepted that ICA presents a reliable method to separate spatially independent source signals from the mixtures. However, ICA does not concern any paradigm information to extract independent components (ICs). Therefore, in this work, we utilize a method of using cICA to better extract ICs of interest from mixtures using prior information of task-related experiments. The details of ICA and cICA are described in the following sections.

Independent component analysis
In general, blind source separation (BSS) is applied to recover original source signals from mixtures. Assuming that the measured signals are presented as xðtÞ ¼ ðx 1 ðtÞ; x 2 ðtÞ; . . . ; x n ðtÞÞ T and the original signals as sðtÞ ¼ ðs 1 ðtÞ; s 2 ðtÞ; . . . ; s m ðtÞÞ T , ICA addresses the BSS problem in a particular situation in which x is a linear mixture of the sources, where A is the mixing matrix with size ðn Â mÞ. The ICA algorithm aims at computing a ðm Â nÞ demixing matrix W ¼ ½w 1 ; w 2 ; . . . ; w m T to recover all ICs from the measurements where yðtÞ ¼ ðy 1 ðtÞ; y 2 ðtÞ; . . . ; y m ðtÞÞ T . Each separated output signal is presented as y i ¼ w T i x or y ¼ w T x. With regards to the central limit theorem, maximizing the non-Gaussianity of y will make it to converge toward one of the ICs close to the original signals. The non-Gaussianity is measured by the negentropy function JðyÞ, expressed as where y Gauss is a Gaussian random variable having the same variance as that of the output signal y, and HðyÞ the entropy of y. Hyvärinen 24 introduced an approximation of negentropy as where is a positive constant that is set as one in our experiments, fðÁÞ is a nonquadratic function, EfÁg is an expectation operator and is a standard Gaussian variable with zero mean and unit variance. Usually, ICA identi¯es as many ICs as the number of observations, and an ordering of the extracted ICs is arbitrary. 25 Paradigm information has been used after ICA to select output components 26 but was not directly involved in the unmixing process of source signals.

Constrained independent component analysis
In this section, we describe the advantage of cICA to extract ICs of interest by de¯ning constraints on the outputs of interest. Mathematically, the basics of cICA are formulated from the ICA algorithm with the main objective to maximize the negentropy term in Eq. (6). Apart from negentropy, additive constraints used to minimize the closeness measurement between the outputs and the references are presented by gðyÞ ¼ "ðy; rÞ À 0 where is the threshold, r the prior information guiding the output signals and "ðy; rÞ the closeness measurement. The typical common form of "ðy; rÞ used in our method is the mean square error (MSE) "ðy; rÞ ¼ Efðy À rÞ 2 g.
The optimization formulation of cICA is expressed by maximize JðyÞ ¼ ðEffðyÞg À EffðÞgÞ 2 subject to gðyÞ 0 ; ð7Þ or minimize J ðy : WÞ ¼ ÀJðyÞ where W is a demixing matrix. Assume that the Lagrange multiplier is , the optimization function is rewritten as The explicit form of Gðy : W; Þ is found in Refs. 25 and 27. The Jacobian matrix of LðW; Þ with respect to w is computed by r 2 w L ¼ ðÀEff 00 ðyÞgþ Efg 00 y ðyÞgÞ § xx where ¼ 2ðEffðyÞg À EffðÞgÞ and § xx ¼ Efxx T g is the covariance matrix.
The fast update rule for a one-unit reference derived by maxf0; þ gðyÞg; w w À r 2 w L À1 ðÀEfJ 0 ðyÞx T g T þ r w gðyÞ T Þ; w w=jjwjj; ð10Þ where is the update rate. The parameters and are¯xed as 0.5 in all experiments. The update procedure described in Eq. (10) is repeated until the algorithm converges. [27][28][29] The reference signal is formed by a binary vector that gets the values of ones within an onset duration (stimulus is turned on) and zeros otherwise. The amplitude of the recovered IC will be estimated by the mixing and demixing matrix. Meanwhile, cICA is implemented by MATLAB 2008a. The cICA extraction procedure to acquire the true responses from the measurements using a reference is illustrated in Fig. 2.

Evaluation metrics
Contrast-to-noise ratio (CNR) 23 is used to re°ect the di®erences between the signals during task and rest. The larger the CNR are, the better fNIRS signals are capable of presenting the actual changes of the brain activities. 30 An equation to compute CNR is expressed by where mean(task) and mean(rest) are the averages of the signals within the task duration and rest duration, respectively, var(task) and var(rest) are the standard variation of the signals within the two durations, and varðtaskÞ þ varðrestÞ is used to normalize the values of CNR with regards to the variation of the signals.

Performance analysis of cICA
We conducted both experiments with real data and simulations to validate the e®ects of cICA on taskrelated/nontask related signals as well as on visual/ nonvisual signals at control regions. For experiments, only oxy-Hb signals were considered. For simulations, we studied whether cICA may return false visual evoked responses from nontarget  stimuli. We recorded 15-min data series (900 s) with a single subject. The subject is advised to take rest while the NIRS data were collected. Five synthetic oxy-Hb responses generated by a Gamma function 26 with the time constant of 3 s were randomly added into the data and separated by at least 60 s. We compared the components extracted by cICA between before and after the synthetic responses were added. The folding-average of the extracted components over¯ve trials shown in Fig. 3 revealed that cICA has not returned a false task-related extraction.
Additionally, we designed a paradigm with real data to analyze the di®erences between signals extracted by visual at control regions. Four subjects were involved in this experiment. Con¯gurations of probes to collect seven-channel data of visual cortex areas were presented in Sec. 2.1.3. Besides, we measured two extra channels A and B at control regions as illustrated in Fig. 4. Only two°ickering frequencies of 0 and 10 Hz have been examined to evoke hemodynamic responses of the brain. During experimental sessions, the visual°ickering was randomly turned on and lasted for 60 s. The trials were repeated for six times. Recovered oxy-Hb and deoxy-Hb ICs of one participant by applying cICA on all measured channels including channels of visual cortex and channels of control regions are shown in Figs. 5 and 6. When all channels were used, corresponding spatial map of the extracted components were described to show that active regions belong to the visual cortex of the brain. Average CNR values in comparing hemodynamic responses between 20 s before and 60 s after a stimulus occurred were reported: CNR values of recovered oxy-Hb signals between visual and control regions are 0.81 and 0.43 for 0 Hz°ickering stimulus and 0.58 and 0.21 for 10 Hz°ickering stimulus. We see that cICA performed better in terms of CNR on the visual cortex region where the recovered hemodynamic responses were associated with the target stimuli.

Hemodynamic responses within task and rest duration
Before considering the hemodynamic changes with regards to the modulation depth and°ickering frequency of lights, we examined how the visual cortex acts with and without the presence of visual stimulus. We applied t-test and CNR to validate the responses of the oxy-Hb and deoxy-Hb concentration within the task and the rest duration. To evaluate the e±ciency of utilizing cICA in extracting the actual responses of fNIRS on a block experiment, we compared the results of cICA with CBSI 23 applied on the fNIRS signals. Paired-sample t-test analysis showed that even when there were no improvement (only smoothing was used), the di®erences of hemodynamic responses between task and rest duration were signi¯cant for both oxy-Hb concentration and deoxy-Hb concentration (p < 0:0001) revealing the fNIRS responses to the visual stimulation. When CBSI and cICA were used, the oxy-Hb and deoxy-Hb concentrations between task and rest remained signi¯cantly di®erent (p < 0:0001). However, the responses of fNIRS within the onset duration became larger when cICA and CBSI were applied and cICA achieved the best results in comparison with other techniques in terms of CNR as illustrated in Fig. 7(a).
The folding-average of the oxy-Hb and deoxy-Hb concentration over all frequencies, modulations and subjects are summarized in Fig. 8 Figure 9 illustrates the average hemodynamic responses between the two modulations depths. Marginally, oxy-Hb concentration tended to decrease (8 over 10 subjects) and deoxy-Hb concentration tended to increase when a lower modulation of°ickering stimulus was present.

Interaction of modulation and frequency of°ickering stimuli to hemodynamic responses of visual cortex
Factorial 2 Â 4 ANOVA were used to analyze the interaction of the two modulation depths and four°i ckering frequencies. The main e®ect reported no interaction between modulation levels and frequencies for oxy-Hb concentration ðF ¼ 0:378; p < 0:769Þ and for deoxy-Hb concentration ðF ¼ 1:231; p < 0:297Þ. The estimated marginal means of oxy-Hb and deoxy-Hb concentration are depicted in Fig. 10. We found that oxy-Hb concentration of the

Conclusions and Discussion
This study investigated the hemodynamic responses of the visual cortex of 10 healthy subjects while stimulating their eyes with a light beam modulated sinusoidally. Frequencies were varied in the range of 0 to 15 Hz and the modulation depth in the range of 0.5 to 1. To enhance the quality of the desired fNIRS signals, we used cICA to improve actual NIRS responses from the obtained event-related fNIRS measurements. The prior information integrated in cICA regulated the estimated ICs toward paradigm information. Experimental results have showed that the proposed method was e®ective to acquire task-related components from the NIRS mixtures: with the help of cICA, CNR re°ecting the contrast of hemodynamic concentration between rest and task became larger, showing improvements of cICA when the noises were present in the NIRS signals.
The experiments to determine the variations of the mean values of oxy-Hb and deoxy-Hb concentration within the task duration revealed signi¯cant changes of the hemodynamic responses of the visual cortex from the baseline when the°ickering stimulation was activated. In subsequent studies, we found that there was slight decrease of the oxygenated hemoglobin concentration over four frequencies when the modulation increased. However, the variations of oxy and deoxy-hemoglobin over di®erent stimulus conditions were not statistically signi¯cant.
It is unclear why we obtained such a weak correlation between photic sinusoidal stimulation and the oxy-Hb and deoxy-Hb changes. In the physiological aspect, such a stimulation induces clear neuronal responses using measurement methods such as psychophysics, electrophysiology and fMRI. The fNIRS is a powerful method to monitor hemodynamic responses of the brain but may not be appropriate to monitor deep neural layers to¯nd coupling between neural responses and oxy-Hb and deoxy-Hb concentrations. Bridge 16 used pattern stimulation in combination with temporal stimulation and also found insigni¯cant di®erences of the brain responses to various temporal frequencies of stimuli. He also suggested that it may be due to fNIRS limitations. The issue of weak correlation between photic sinusoidal stimulation and the oxy-Hb and deoxy-Hb changes is still open for discussion and further investigations.