Periodic Reporting for period 1 - ThalamicAttention (The role of human thalamus in selective attention via novel denoising applied to magnetoencephalography)
Reporting period: 2022-11-01 to 2024-10-31
Objective 1. Source localization method development through improving volume conduction estimates (forward model) and denoising signal with SVD during beamforming (inverse model).
ThalamicAttention set out to achieve MEG source localization methods development via improvements of the inverse model (i.e. algorithms projecting the MEG sensor signal to source space) and the forward model (i.e. modelling the signal from source space to MEG sensors). The forward model is an essential ingredient of the inverse model, thus improvements of both contribute to obtaining more accurate source localization results with MEG.
At the heart of the inverse model improvement is the idea of using a modified version of conventional beamforming (i.e. eigenspace beamforming) algorithm that was designed to restrict source localization of the signal to Principal Component Analysis (PCA) components that are not
considered noise. An obstacle of using this method was being able to separate actual signal and noise components. To this purpose I used a denoising method that can threshold PCA or SVD components of covariance matrices based on the Marchenko-Pastur distribution (MP-denoising) which in turn can be fed into the modified (i.e. eigenspace) beamformer. To account for the high dimensional nature of typical electrophysiological data, I also set out to implement a version of MP-denoising, tensor SVD MP-denoising whereby the MP-denosing is carried out on a 3D (or higher) dimensional matrix (i.e. tensor).
Objective 2. Tapping into a range of thalamic signals (ASSR, VSSR and α-band) through improved SNR via the application of the developed source localization method with MEG data.
This objective was aimed at tapping into thalamic signals in recorded experimental MEG data. The experiment was designed specifically to provide data as closely to achieving a ground truth signal as possible. The experiment used a set frequency auditory (43Hz) and visual (24Hz) stimulation which is well known to result in a frequency following response at the same frequency in the respective sensory cortices (thus establishing ground truth of primary signal processing). Importantly, it is also known to have a similar frequency following response in sensory thalamic nuclei. In addition, the visual and the auditory sensory pathways are well mapped out from ear/retina to sensory areas with established time delays across the pathways. As a result, we can have very clear expectations regarding area, frequency and time delay of the signal.
Objective 3. Examining cross-modal attention effects in the thalamus on ASSR, VSSR and α-band signal via the application of the developed source localization method with MEG data.
The experimental data also served the purpose of assessing if attention directed towards auditory versus visual stimulation results in detectable differences at any particular brain region and especially in the thalamus as measured with MEG. The conceptual aim was to address whether attention filtering across modalities already starts at the level of the thalamus and to find
electrophysiological correlates of such cross-modal attention allocation across the brain in general.
ThalamicAttention set out to achieve MEG source localization methods development via improvements of the inverse model (i.e. algorithms projecting the MEG sensor signal to source space) and the forward model (i.e. modelling the signal from source space to MEG sensors). The forward model is an essential ingredient of the inverse model, thus improvements of both contribute to obtaining more accurate source localization results with MEG.
At the heart of the inverse model improvement is the idea of using a modified version of conventional beamforming (i.e. eigenspace beamforming) algorithm that was designed to restrict source localization of the signal to Principal Component Analysis (PCA) components that are not
considered noise. An obstacle of using this method was being able to separate actual signal and noise components. To this purpose I used a denoising method that can threshold PCA or SVD components of covariance matrices based on the Marchenko-Pastur distribution (MP-denoising) which in turn can be fed into the modified (i.e. eigenspace) beamformer. To account for the high dimensional nature of typical electrophysiological data, I also set out to implement a version of MP-denoising, tensor SVD MP-denoising whereby the MP-denosing is carried out on a 3D (or higher) dimensional matrix (i.e. tensor).
Objective 2. Tapping into a range of thalamic signals (ASSR, VSSR and α-band) through improved SNR via the application of the developed source localization method with MEG data.
This objective was aimed at tapping into thalamic signals in recorded experimental MEG data. The experiment was designed specifically to provide data as closely to achieving a ground truth signal as possible. The experiment used a set frequency auditory (43Hz) and visual (24Hz) stimulation which is well known to result in a frequency following response at the same frequency in the respective sensory cortices (thus establishing ground truth of primary signal processing). Importantly, it is also known to have a similar frequency following response in sensory thalamic nuclei. In addition, the visual and the auditory sensory pathways are well mapped out from ear/retina to sensory areas with established time delays across the pathways. As a result, we can have very clear expectations regarding area, frequency and time delay of the signal.
Objective 3. Examining cross-modal attention effects in the thalamus on ASSR, VSSR and α-band signal via the application of the developed source localization method with MEG data.
The experimental data also served the purpose of assessing if attention directed towards auditory versus visual stimulation results in detectable differences at any particular brain region and especially in the thalamus as measured with MEG. The conceptual aim was to address whether attention filtering across modalities already starts at the level of the thalamus and to find
electrophysiological correlates of such cross-modal attention allocation across the brain in general.
Objective 1.
As part of the project the necessary software implementations including the crucial tensor SVD MP-denoising has been completed and the denoising method had been incorporated into the eigenspace beamforming script in the Matlab-based Fieldtrip toolbox meeting the project schedule. Following the implementation the method’s performance on simulated and actual data has been carefully tested and assessed, and it was concluded that the gain of utilizing the method did not meet expectations. In effect only a small noise reduction was achieved on real data, complicated by additional drawbacks of the method which then made us reach the conclusion that the pursuit of a wider use is not warranted (see more details at WP1). This part of the project was perfectly on schedule up to the implementation and analysis stage. However, further planned measures i.e. publication of the method and tutorials of its use, were not taken and the objective was completed without further action. As an alternative measure to achieve improvements on the inverse model, I sought out to implement a Bayesian beamformer for use with magnetometers and planar gradiometers, which has the promise of overcoming one of the most significant problems of beamforming, degraded performance with simultaneous and correlated sources. This work started halfway through the project and is still under way. Improvement of the forward model was aimed to be achieved by obtaining better conductivity
estimates, an important component of accurate forward modelling. This was approached via two ways: (a) updating the currently used standards for white and grey matter conductivities through analysing conductivity measurements of live human tissue samples, (b) implementing a combinatorial estimation of optimal skull, scalp, grey matter, white matter and cerebrospinal fluid conductivity. The updating of currently used brain tissue conductivity estimates had to be abandoned at the very start of the project due to the sudden passing of a key collaborator, Prof Capogna, who was
essential and irreplaceable to the access to live tissue samples and the measurements of the samples. Since it was not feasible to acquire samples and measurements from other institutions, I took an alternative measure with the specific aim of improving the conductivity estimates through different means. I set out to run the cross-modal attention task - designed for MEG – also as an fMRI task. The rationale behind is that using the BOLD signal as an independent marker of brain regions primarily involved in processing the visual and the auditory stimuli, I can use the fMRI results to identify the most optimal conductivity estimates for MEG source analysis. Theoretically, the method could be used as a generic method, irrespective of the task by simply running short visual and auditory stimulation (i.e. localizer sequences) in both MEG/EEG and fMRI. To examine the feasibility of this method I implemented and ran the cross- modal attention task during fMRI scanning with the same participants who performed the identical task in the MEG. The analysis of the results is currently ongoing. Improving forward modelling via the optimization of conductivity parameters entered into computations was set out through combinatorial approximation of best conductivity estimates of layer conductivities (skull, scalp, white matter, grey matter, CSF). Essentially, the method suggests finding the most optimal conductivity estimates through systematically changing layer conductivities and assessing the source localization results. As part of the objective the method has been implemented with advanced forward modelling with Boundary Element Method and Finite Element Method , however further testing of the method on the full experimental study has been shifted due to delay of the experimental data collection. The assessment of the method’s performance is currently ongoing.
Objective 2.
SNR improvement was approached via developments of source localization through optimizing forward and inverse modelling (see Objective 1). The experiment crucial for Objective 2 has been implemented and was conducted, the recorded data have been preprocessed and analysis has been ongoing/partially completed. Even with standard source localization methods, thalamic activity has been detected at the visual stimulation frequency, but so far not at the auditory stimulation frequency. The mismatch is likely due to the much stronger brain response to visual stimulation than to auditory stimulation, at least in sensory cortical areas. To my knowledge this is the first clear demonstration of thalamic signals in response to visual (or any other) brain stimulation via non-invasive electrophysiological methods. If the ongoing optimization of forward modelling and modified beamforming will achieve SNR improvement and better detection of visual and auditory signals in the thalamus remains to be seen as analysis and optimization of the methods as discussed above (Objective 1) is largely ongoing. Note that the assessment of MP-denoising has been completed as the method did not show significant enough SNR improvement at sensor or source level, as well as, specifically detecting thalamic signals (see Objective 1).
Objective 3.
Achieving the objective required the collection of the above mentioned experimental data and the preprocessing and analysis of the collected data. This has been fully achieved and whilst no cross-modal attention effect was identified in the thalamus, a series of electrophysiological correlates have been detected in various (primarily posterior) brain areas. The writing up of the results of the MEG findings is underway currently aimed at being submitted to a high-impact open-access journal. In addition, the collected fMRI data proved to be highly important at identifying effects largely complementary to the MEG findings and thus contributing to refining our models of cross-modal attention filtering. The first draft of the planned paper disseminating these fMRI findings is being prepared. Currently, the MEG and the fMRI results are being prepared as separate publications due to the large volume of findings with both methods.
As part of the project the necessary software implementations including the crucial tensor SVD MP-denoising has been completed and the denoising method had been incorporated into the eigenspace beamforming script in the Matlab-based Fieldtrip toolbox meeting the project schedule. Following the implementation the method’s performance on simulated and actual data has been carefully tested and assessed, and it was concluded that the gain of utilizing the method did not meet expectations. In effect only a small noise reduction was achieved on real data, complicated by additional drawbacks of the method which then made us reach the conclusion that the pursuit of a wider use is not warranted (see more details at WP1). This part of the project was perfectly on schedule up to the implementation and analysis stage. However, further planned measures i.e. publication of the method and tutorials of its use, were not taken and the objective was completed without further action. As an alternative measure to achieve improvements on the inverse model, I sought out to implement a Bayesian beamformer for use with magnetometers and planar gradiometers, which has the promise of overcoming one of the most significant problems of beamforming, degraded performance with simultaneous and correlated sources. This work started halfway through the project and is still under way. Improvement of the forward model was aimed to be achieved by obtaining better conductivity
estimates, an important component of accurate forward modelling. This was approached via two ways: (a) updating the currently used standards for white and grey matter conductivities through analysing conductivity measurements of live human tissue samples, (b) implementing a combinatorial estimation of optimal skull, scalp, grey matter, white matter and cerebrospinal fluid conductivity. The updating of currently used brain tissue conductivity estimates had to be abandoned at the very start of the project due to the sudden passing of a key collaborator, Prof Capogna, who was
essential and irreplaceable to the access to live tissue samples and the measurements of the samples. Since it was not feasible to acquire samples and measurements from other institutions, I took an alternative measure with the specific aim of improving the conductivity estimates through different means. I set out to run the cross-modal attention task - designed for MEG – also as an fMRI task. The rationale behind is that using the BOLD signal as an independent marker of brain regions primarily involved in processing the visual and the auditory stimuli, I can use the fMRI results to identify the most optimal conductivity estimates for MEG source analysis. Theoretically, the method could be used as a generic method, irrespective of the task by simply running short visual and auditory stimulation (i.e. localizer sequences) in both MEG/EEG and fMRI. To examine the feasibility of this method I implemented and ran the cross- modal attention task during fMRI scanning with the same participants who performed the identical task in the MEG. The analysis of the results is currently ongoing. Improving forward modelling via the optimization of conductivity parameters entered into computations was set out through combinatorial approximation of best conductivity estimates of layer conductivities (skull, scalp, white matter, grey matter, CSF). Essentially, the method suggests finding the most optimal conductivity estimates through systematically changing layer conductivities and assessing the source localization results. As part of the objective the method has been implemented with advanced forward modelling with Boundary Element Method and Finite Element Method , however further testing of the method on the full experimental study has been shifted due to delay of the experimental data collection. The assessment of the method’s performance is currently ongoing.
Objective 2.
SNR improvement was approached via developments of source localization through optimizing forward and inverse modelling (see Objective 1). The experiment crucial for Objective 2 has been implemented and was conducted, the recorded data have been preprocessed and analysis has been ongoing/partially completed. Even with standard source localization methods, thalamic activity has been detected at the visual stimulation frequency, but so far not at the auditory stimulation frequency. The mismatch is likely due to the much stronger brain response to visual stimulation than to auditory stimulation, at least in sensory cortical areas. To my knowledge this is the first clear demonstration of thalamic signals in response to visual (or any other) brain stimulation via non-invasive electrophysiological methods. If the ongoing optimization of forward modelling and modified beamforming will achieve SNR improvement and better detection of visual and auditory signals in the thalamus remains to be seen as analysis and optimization of the methods as discussed above (Objective 1) is largely ongoing. Note that the assessment of MP-denoising has been completed as the method did not show significant enough SNR improvement at sensor or source level, as well as, specifically detecting thalamic signals (see Objective 1).
Objective 3.
Achieving the objective required the collection of the above mentioned experimental data and the preprocessing and analysis of the collected data. This has been fully achieved and whilst no cross-modal attention effect was identified in the thalamus, a series of electrophysiological correlates have been detected in various (primarily posterior) brain areas. The writing up of the results of the MEG findings is underway currently aimed at being submitted to a high-impact open-access journal. In addition, the collected fMRI data proved to be highly important at identifying effects largely complementary to the MEG findings and thus contributing to refining our models of cross-modal attention filtering. The first draft of the planned paper disseminating these fMRI findings is being prepared. Currently, the MEG and the fMRI results are being prepared as separate publications due to the large volume of findings with both methods.
The project revealed a number of findings offering novel insights into cross-modal attention processes. Crucially, it showed a large posterior power difference when attention was directed to the auditory stimulation stream relative to the visual stream. The effect was largest in the low frequency range and in particular around 14-22Hz (beta band), showing significantly higher power in the attend auditory condition. The source localization indicates a large posterior cluster that includes a number of brain regions primarily involved in visual processing. The effect was particularly strong in this beta range, but was also present in very low frequencies (2-8Hz). This finding is surprising for a number of reasons. First and foremost, beta frequency has been associated with attention processes before, but primarily as a bottom up stimulus driven, attention capturing effect. Here however, it clearly represents a top-down selective attention process.
Second, the low frequency power effect is in line with previous assumptions and findings of low frequency (primarily delta) power decrease due to switching from a highly synchronous resting state to a more desynchronized state in processing relevant areas. Lastly, the lack of any alpha frequency effect (8-12Hz) is rather unexpected. Alpha desynchronization accompanying attention modulation and also visual stimulus processing is a well established phenomenon. Hence, the lack of alpha power modulation is a puzzling finding.
In addition, the frequency-following brain response, at the same frequency as the induced visual and auditory signal has been reported to show cross-modal attention modulation, although the effect being stronger with EEG and extremely weak with MEG. The results of this project, do not fully support these findings. No auditory brain response (43Hz) and no visual brain response (24Hz) attention modulation have been found at the primary frequencies. This is despite the fact that the 24Hz signal is extremely consistent and strong across participants, indicating that SNR cannot account for the lack of effect. The only effect that emerged was at 48Hz (higher harmonics of the induced 24Hz) and only when analyzing the evoked power of the signal (i.e. taking the average of the signal across trials and than calculating the power). Whilst this is a clear indication of some attention modulation of the induced signal, it is clearly too small to account for cross-modal attention filtering in general.
An important further finding is that the visual 24Hz signal can be detected even with the most conventional and least detailed forward and standard DICS beamforming. Similar results were found with BEM. Although given the large spatial posterior cluster that shows significant power increase at the stimulation frequency relative to baseline, it does not completely allow for ruling out spatial smear from other regions, the characteristics of the signal indicate that likely the thalamic signal is captured. Even though there is no attention modulation of the signal in the thalamus, this is still an important finding given that the thalamic signal is thought to be extremely difficult to tap into with MEG.
Based on the spatial profile of the MEG results it is tempting to conclude that visual filtering in cross-modal attention is likely filtered through low frequency beta mechanisms potentially covering a large part of the posterior section of the brain. At the same time, there is little we can say about filtering/facilitating auditory processing and primary and secondary auditory areas may not be involved in attention processes at all. This interpretation however is likely incorrect as the fMRI results tell a different story. the results clearly show double dissociations
of the BOLD signal response due to attention directed towards the auditory versus the visual stream. The results indicate not only early visual cortical but also higher visual areas showing significantly stronger BOLD signal responses when attention is directed towards the visual stream. Importantly, significantly higher BOLD activity is present in early auditory areas and somewhat unexpectedly in the Precuneus, a medial posterior area rarely thought to be involved in auditory processing. An alternative interpretation is that the higher BOLD signal of the Precuneus is primarily aiding the filtering of the visual signal (i.e. not specifically associated with facilitation of the auditory signal but the filtering of the visual signal instead). There is currently no way of unequivocally determining the role of the Precuneus activity, but other areas indicated by BOLD signal change likely represent the facilitation/inhibition of the sensory signal in the respective sensory area. With respect to the thalamus, I found BOLD response increase in the thalamus during perceptual stimulation (auditory and visual) as compared to a stimulation silent baseline period; but with no attention effect.
Second, the low frequency power effect is in line with previous assumptions and findings of low frequency (primarily delta) power decrease due to switching from a highly synchronous resting state to a more desynchronized state in processing relevant areas. Lastly, the lack of any alpha frequency effect (8-12Hz) is rather unexpected. Alpha desynchronization accompanying attention modulation and also visual stimulus processing is a well established phenomenon. Hence, the lack of alpha power modulation is a puzzling finding.
In addition, the frequency-following brain response, at the same frequency as the induced visual and auditory signal has been reported to show cross-modal attention modulation, although the effect being stronger with EEG and extremely weak with MEG. The results of this project, do not fully support these findings. No auditory brain response (43Hz) and no visual brain response (24Hz) attention modulation have been found at the primary frequencies. This is despite the fact that the 24Hz signal is extremely consistent and strong across participants, indicating that SNR cannot account for the lack of effect. The only effect that emerged was at 48Hz (higher harmonics of the induced 24Hz) and only when analyzing the evoked power of the signal (i.e. taking the average of the signal across trials and than calculating the power). Whilst this is a clear indication of some attention modulation of the induced signal, it is clearly too small to account for cross-modal attention filtering in general.
An important further finding is that the visual 24Hz signal can be detected even with the most conventional and least detailed forward and standard DICS beamforming. Similar results were found with BEM. Although given the large spatial posterior cluster that shows significant power increase at the stimulation frequency relative to baseline, it does not completely allow for ruling out spatial smear from other regions, the characteristics of the signal indicate that likely the thalamic signal is captured. Even though there is no attention modulation of the signal in the thalamus, this is still an important finding given that the thalamic signal is thought to be extremely difficult to tap into with MEG.
Based on the spatial profile of the MEG results it is tempting to conclude that visual filtering in cross-modal attention is likely filtered through low frequency beta mechanisms potentially covering a large part of the posterior section of the brain. At the same time, there is little we can say about filtering/facilitating auditory processing and primary and secondary auditory areas may not be involved in attention processes at all. This interpretation however is likely incorrect as the fMRI results tell a different story. the results clearly show double dissociations
of the BOLD signal response due to attention directed towards the auditory versus the visual stream. The results indicate not only early visual cortical but also higher visual areas showing significantly stronger BOLD signal responses when attention is directed towards the visual stream. Importantly, significantly higher BOLD activity is present in early auditory areas and somewhat unexpectedly in the Precuneus, a medial posterior area rarely thought to be involved in auditory processing. An alternative interpretation is that the higher BOLD signal of the Precuneus is primarily aiding the filtering of the visual signal (i.e. not specifically associated with facilitation of the auditory signal but the filtering of the visual signal instead). There is currently no way of unequivocally determining the role of the Precuneus activity, but other areas indicated by BOLD signal change likely represent the facilitation/inhibition of the sensory signal in the respective sensory area. With respect to the thalamus, I found BOLD response increase in the thalamus during perceptual stimulation (auditory and visual) as compared to a stimulation silent baseline period; but with no attention effect.