Thomas Knösche - Academia.edu (original) (raw)

Papers by Thomas Knösche

Human brain mapping, Jan 14, 2015

In this study, we used invasive tracing to evaluate white matter tractography methods based on ex... more In this study, we used invasive tracing to evaluate white matter tractography methods based on ex vivo diffusion-weighted magnetic resonance imaging (dwMRI) data. A representative selection of tractography methods were compared to manganese tracing on a voxel-wise basis, and a more qualitative assessment examined whether, and to what extent, certain fiber tracts and gray matter targets were reached. While the voxel-wise agreement was very limited, qualitative assessment revealed that tractography is capable of finding the major fiber tracts, although there were some differences between the methods. However, false positive connections were very common and, in particular, we discovered that it is not possible to achieve high sensitivity (i.e., few false negatives) and high specificity (i.e., few false positives) at the same time. Closer inspection of the results led to the conclusion that these problems mainly originate from regions with complex fiber arrangements or high curvature an...

Background / Purpose: In this work, we evaluate several standard visualization techniques in thei... more Background / Purpose: In this work, we evaluate several standard visualization techniques in their applicability on electrical fields from EEG and tDCS in the human brain. Main conclusion: Most applications profit from visualization, but the used visualization technique is heavily use-case dependent and needs to be chosen with care.

Biomedizinische Technik/Biomedical Engineering

NeuroImage, 2009

Diffusion Tensor Imaging (DTI) is generally used for the analysis of white matter tissue anisotro... more Diffusion Tensor Imaging (DTI) is generally used for the analysis of white matter tissue anisotropy while cortical gray matter does not show clear anisotropy in most diffusion imaging acquisitions. However, microscopic analysis of the cortex shows variable microscructure in different functioanatomical areas. Cortical areas with large pyramidal cells, which are oriented in normal direction with respect to the folded cortical surface, as in the motor cortex, might show radial anisotropy in high resolution DTI scans. The dominant direction should differ from other cortical areas with less dominant cell structure as the primary somato-sensory cortex. In this study we analyze the cortical anisotropy with a new radiality coefficient and test, if the principal diffusion direction depends on the type of the cortex.

The cortical network subserving language processing is likely to exhibit a high spatial and tempo... more The cortical network subserving language processing is likely to exhibit a high spatial and temporal complexity. Studies using brain imaging methods, like fMRI or PET, succeeded in identifying a number of brain structures that seem to contribute to the processing of syntactic structures, while their dynamic interaction remains unclear due to the low temporal resolution of the methods. On the other hand, ERP studies have revealed a great deal of the temporal dimension of language processing without being able to provide more than very coarse information on the localisation of the underlying generators. MEG has a temporal resolution similar to EEG combined with a better spatial resolution. In this paper, Brain Surface Current Density (BSCD) mapping in a standard brain model was used to identify statistically significant differences between the activity of certain brain regions due to syntactically correct and incorrect auditory language input. The results show that the activity in the first 600 ms after violation onset is mainly concentrated in the temporal cortex and the adjacent frontal and parietal areas of both hemispheres. The statistical analysis reveals significantly different activity mainly in both frontal and temporal cortices. For longer latencies above 250 ms, the differential activity is more prominent in the right hemisphere. These findings confirm other recent results that suggest right hemisphere involvement in auditory language processing. One interpretation might be that right hemisphere regions play an important role in repair and re-analysis processes in order to free the specialised left hemisphere language areas for processing further input.

Lecture Notes in Computer Science, 2007

This paper presents a method to reduce time complexity of the computation of higher-order tensor ... more This paper presents a method to reduce time complexity of the computation of higher-order tensor lines. The method can be applied to higher-order tensors and the spherical harmonics representation, both widely used in medical imaging. It is based on a gradient descend technique and integrates well into fiber tracking algorithms. Furthermore, the method improves the angular resolution in contrast to discrete sampling methods which is especially important to tractography, since there, small errors accumulate fast and make the result unusable. Our implementation does not interpolate derived directions but works directly on the interpolated tensor information. The specific contribution of this paper is a fast algorithm for tracking lines tensor fields of arbitrary order that increases angular resolution compared to previous approaches.

Common approaches for distributed source reconstruction are the minimum norm (MN), weighted MN (W... more Common approaches for distributed source reconstruction are the minimum norm (MN), weighted MN (WMN) and LORETA method. While MN weights sources equally, WMN gives more impact to deep sources, which implies an assumption about source variances. LORETA additionally imposes source covariance using a Laplacian operator as a smoothness criterion. We propose a new method, LineLORETA, that allows incorporating individual functio-anatomical information with the Laplacian, e. g. functional borders, leading to a functio-anatomically inspired source covariance estimate. Such priors can be taken from general functio-anatomical knowledge applied to the individual cortical sheet, or from similarity measures as derived from, e.g., fMRI. In LineLORETA, the Laplacian is modified by cutting the geometrical neighborhood at functio-anatomical boundaries. A valid source covariance estimate is achieved by applying normalization to the inverted Laplacian. A regularization parameter controls the impact of...

Die Messung neuroelektromagnetischer Signale (EEG/MEG) bietet gegenüber anderen Modalitäten eine ... more Die Messung neuroelektromagnetischer Signale (EEG/MEG) bietet gegenüber anderen Modalitäten eine hervorragende Zeitauflösung und damit prinzipiell die Möglichkeit, Gehirnaktivität in unmittelbarer Reaktion auf Stimuli mithilfe von Quellenrekonstruktionsalgorithmen zu analysieren. Wir stellen ein Konzept für ein System vor, welches zur Ausführung der notwendigen Signalaufbereitungsowie zur Durchführung verteilter Quellenanalysen zur Laufzeit in der Lage ist und damit Online-Analysen ermöglicht. Parameter lassen sich zur Optimierung der Signalqualität zur Laufzeit anpassen.

NeuroImage, Jan 15, 2015

The results of brain connectivity analysis using reconstructed source time courses derived from E... more The results of brain connectivity analysis using reconstructed source time courses derived from EEG and MEG data depend on a number of algorithmic choices. While previous studies have investigated the influence of the choice of source estimation method or connectivity measure, the effects of the head modeling errors or simplifications have not been studied sufficiently. In the present simulation study, we investigated the influence of particular properties of the head model on the reconstructed source time courses as well as on source connectivity analysis in EEG and MEG. Therefore, we constructed a realistic head model and applied the finite element method to solve the EEG and MEG forward problems. We considered the distinction between white and gray matter, the distinction between compact and spongy bone, the inclusion of a cerebrospinal fluid (CSF) compartment, and the reduction to a simple 3-layer model comprising only the skin, skull, and brain. Source time courses were reconst...

Electroencephalography and clinical neurophysiology, 1998

The application of surface laplacian and linear estimation methods to single trial EEG data was s... more The application of surface laplacian and linear estimation methods to single trial EEG data was studied. EEG was recorded in 3 subjects during voluntary, self-paced extensions and flexions of the index finger. In each subject a post-movement beta synchronisation was found in specific frequency bands. The surface laplacian estimates were calculated using spherical splines and cortical current distributions were constructed using the linear estimation method. Both methods yield similar results and reveal a maximal event-related synchronisation over the left sensorimotor area approximately 500-750 ms after termination of movement.

NeuroImage, Jan 15, 2015

We present an MEG source reconstruction method that simultaneously reconstructs source amplitudes... more We present an MEG source reconstruction method that simultaneously reconstructs source amplitudes and identifies source interactions across the whole brain. In the proposed method, a full multivariate autoregressive (MAR) model formulates directed interactions (i.e., effective connectivity) between sources. The MAR coefficients (the entries of the MAR matrix) are constrained by the prior knowledge of whole-brain anatomical networks inferred from diffusion MRI. Moreover, to increase the accuracy and robustness of our method, we apply an fMRI prior on the spatial activity patterns and a sparse prior on the MAR coefficients. The observation process of MEG data, the source dynamics, and a series of the priors are combined into a Bayesian framework using a state-space representation. The parameters, such as the source amplitudes and the MAR coefficients, are jointly estimated from a variational Bayesian learning algorithm. By formulating the source dynamics in the context of MEG source r...

Biomedical Engineering / Biomedizinische Technik, 2012

The reconstruction of distributed sources from EEG/MEG data requires additional assumptions on th... more The reconstruction of distributed sources from EEG/MEG data requires additional assumptions on the nature of the underlying sources. While the solution space is often restricted to the cortical sheet, the LORETA method [1,2] additionally imposes covariances between sources based on a smoothness constraint. Functional information, e.g. known from fMRI experiments, can be encoded in the main diagonal elements of a source weighting matrix to bias the solution. Alternatively, priors on the similarity between neighbouring reconstruction points can be used. While it is difficult to provide such a similarity measure for the whole cortex, it is sensible to assume similarity between neighbouring sources and to lift it were a change of functioanatomical properties of the tissue is expected, e.g. at major sulcus lines. We present a new method, LineLORETA, which uses a functio-anatomical inspired source covariance matrix estimated from the combination of similarity information with a general smoothness assumption.

Biomedical Engineering / Biomedizinische Technik, 2012

A System for the Online Reconstruction of Distributed Sources from EEG and MEG data Neuroelectrom... more A System for the Online Reconstruction of Distributed Sources from EEG and MEG data Neuroelectromagnetic signals given by EEG/MEG measurements provide an excellent time resolution of underlying brain processes. A common method to localize brain activity is distributed source reconstruction, particularly in case of analyzing event related potentials. Such analysis requires several preprocessing steps, e.g. artefact detection and correction, filtering, epoch separation and averaging, and is usually performed offline. Online source localization would provide the visualization of brain activity during the measurement, which is interesting for both medical and research applications. We developed a concept that allows to set up and tune an online signal processing chain, including distributed source localization on individual head models. The implemenation is based on OpenWalnut, a software dedicated to multimodal brain visualization [1,2]. The capabilites of online processing were tested based on streaming a previously recorded dataset and measuring the execution times.

2013 IEEE 7th International Conference on Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2013

ABSTRACT Electroencephalography (EEG) and magnetoencephalography (MEG) provide insight into neuro... more ABSTRACT Electroencephalography (EEG) and magnetoencephalography (MEG) provide insight into neuronal processes in the brain in a real-time scale. Brain activity can be modeled in terms of a source distribution found by solving the bioelectromagnetic inverse problem, e.g. using linear source reconstruction methods. Such methods are particularly suitable to be used on modern highly parallel processing systems, such as widely available graphic processing units (GPUs). The utilization of these capabilities paves the way for online neuroelectromagnetic source imaging. We present a system that, according to its modular scheme, can be configured in a very flexible way using graphical building blocks. It allows to use different preprocessing algorithms together with a linear source reconstruction method. The algorithms use both CPU and GPU resources.

International Review of Neurobiology, 2005

Human brain mapping, Jan 14, 2015

In this study, we used invasive tracing to evaluate white matter tractography methods based on ex... more In this study, we used invasive tracing to evaluate white matter tractography methods based on ex vivo diffusion-weighted magnetic resonance imaging (dwMRI) data. A representative selection of tractography methods were compared to manganese tracing on a voxel-wise basis, and a more qualitative assessment examined whether, and to what extent, certain fiber tracts and gray matter targets were reached. While the voxel-wise agreement was very limited, qualitative assessment revealed that tractography is capable of finding the major fiber tracts, although there were some differences between the methods. However, false positive connections were very common and, in particular, we discovered that it is not possible to achieve high sensitivity (i.e., few false negatives) and high specificity (i.e., few false positives) at the same time. Closer inspection of the results led to the conclusion that these problems mainly originate from regions with complex fiber arrangements or high curvature an...

Background / Purpose: In this work, we evaluate several standard visualization techniques in thei... more Background / Purpose: In this work, we evaluate several standard visualization techniques in their applicability on electrical fields from EEG and tDCS in the human brain. Main conclusion: Most applications profit from visualization, but the used visualization technique is heavily use-case dependent and needs to be chosen with care.

Biomedizinische Technik/Biomedical Engineering

NeuroImage, 2009

Diffusion Tensor Imaging (DTI) is generally used for the analysis of white matter tissue anisotro... more Diffusion Tensor Imaging (DTI) is generally used for the analysis of white matter tissue anisotropy while cortical gray matter does not show clear anisotropy in most diffusion imaging acquisitions. However, microscopic analysis of the cortex shows variable microscructure in different functioanatomical areas. Cortical areas with large pyramidal cells, which are oriented in normal direction with respect to the folded cortical surface, as in the motor cortex, might show radial anisotropy in high resolution DTI scans. The dominant direction should differ from other cortical areas with less dominant cell structure as the primary somato-sensory cortex. In this study we analyze the cortical anisotropy with a new radiality coefficient and test, if the principal diffusion direction depends on the type of the cortex.

The cortical network subserving language processing is likely to exhibit a high spatial and tempo... more The cortical network subserving language processing is likely to exhibit a high spatial and temporal complexity. Studies using brain imaging methods, like fMRI or PET, succeeded in identifying a number of brain structures that seem to contribute to the processing of syntactic structures, while their dynamic interaction remains unclear due to the low temporal resolution of the methods. On the other hand, ERP studies have revealed a great deal of the temporal dimension of language processing without being able to provide more than very coarse information on the localisation of the underlying generators. MEG has a temporal resolution similar to EEG combined with a better spatial resolution. In this paper, Brain Surface Current Density (BSCD) mapping in a standard brain model was used to identify statistically significant differences between the activity of certain brain regions due to syntactically correct and incorrect auditory language input. The results show that the activity in the first 600 ms after violation onset is mainly concentrated in the temporal cortex and the adjacent frontal and parietal areas of both hemispheres. The statistical analysis reveals significantly different activity mainly in both frontal and temporal cortices. For longer latencies above 250 ms, the differential activity is more prominent in the right hemisphere. These findings confirm other recent results that suggest right hemisphere involvement in auditory language processing. One interpretation might be that right hemisphere regions play an important role in repair and re-analysis processes in order to free the specialised left hemisphere language areas for processing further input.

Lecture Notes in Computer Science, 2007

This paper presents a method to reduce time complexity of the computation of higher-order tensor ... more This paper presents a method to reduce time complexity of the computation of higher-order tensor lines. The method can be applied to higher-order tensors and the spherical harmonics representation, both widely used in medical imaging. It is based on a gradient descend technique and integrates well into fiber tracking algorithms. Furthermore, the method improves the angular resolution in contrast to discrete sampling methods which is especially important to tractography, since there, small errors accumulate fast and make the result unusable. Our implementation does not interpolate derived directions but works directly on the interpolated tensor information. The specific contribution of this paper is a fast algorithm for tracking lines tensor fields of arbitrary order that increases angular resolution compared to previous approaches.

Common approaches for distributed source reconstruction are the minimum norm (MN), weighted MN (W... more Common approaches for distributed source reconstruction are the minimum norm (MN), weighted MN (WMN) and LORETA method. While MN weights sources equally, WMN gives more impact to deep sources, which implies an assumption about source variances. LORETA additionally imposes source covariance using a Laplacian operator as a smoothness criterion. We propose a new method, LineLORETA, that allows incorporating individual functio-anatomical information with the Laplacian, e. g. functional borders, leading to a functio-anatomically inspired source covariance estimate. Such priors can be taken from general functio-anatomical knowledge applied to the individual cortical sheet, or from similarity measures as derived from, e.g., fMRI. In LineLORETA, the Laplacian is modified by cutting the geometrical neighborhood at functio-anatomical boundaries. A valid source covariance estimate is achieved by applying normalization to the inverted Laplacian. A regularization parameter controls the impact of...

Die Messung neuroelektromagnetischer Signale (EEG/MEG) bietet gegenüber anderen Modalitäten eine ... more Die Messung neuroelektromagnetischer Signale (EEG/MEG) bietet gegenüber anderen Modalitäten eine hervorragende Zeitauflösung und damit prinzipiell die Möglichkeit, Gehirnaktivität in unmittelbarer Reaktion auf Stimuli mithilfe von Quellenrekonstruktionsalgorithmen zu analysieren. Wir stellen ein Konzept für ein System vor, welches zur Ausführung der notwendigen Signalaufbereitungsowie zur Durchführung verteilter Quellenanalysen zur Laufzeit in der Lage ist und damit Online-Analysen ermöglicht. Parameter lassen sich zur Optimierung der Signalqualität zur Laufzeit anpassen.

NeuroImage, Jan 15, 2015

The results of brain connectivity analysis using reconstructed source time courses derived from E... more The results of brain connectivity analysis using reconstructed source time courses derived from EEG and MEG data depend on a number of algorithmic choices. While previous studies have investigated the influence of the choice of source estimation method or connectivity measure, the effects of the head modeling errors or simplifications have not been studied sufficiently. In the present simulation study, we investigated the influence of particular properties of the head model on the reconstructed source time courses as well as on source connectivity analysis in EEG and MEG. Therefore, we constructed a realistic head model and applied the finite element method to solve the EEG and MEG forward problems. We considered the distinction between white and gray matter, the distinction between compact and spongy bone, the inclusion of a cerebrospinal fluid (CSF) compartment, and the reduction to a simple 3-layer model comprising only the skin, skull, and brain. Source time courses were reconst...

Electroencephalography and clinical neurophysiology, 1998

The application of surface laplacian and linear estimation methods to single trial EEG data was s... more The application of surface laplacian and linear estimation methods to single trial EEG data was studied. EEG was recorded in 3 subjects during voluntary, self-paced extensions and flexions of the index finger. In each subject a post-movement beta synchronisation was found in specific frequency bands. The surface laplacian estimates were calculated using spherical splines and cortical current distributions were constructed using the linear estimation method. Both methods yield similar results and reveal a maximal event-related synchronisation over the left sensorimotor area approximately 500-750 ms after termination of movement.

NeuroImage, Jan 15, 2015

We present an MEG source reconstruction method that simultaneously reconstructs source amplitudes... more We present an MEG source reconstruction method that simultaneously reconstructs source amplitudes and identifies source interactions across the whole brain. In the proposed method, a full multivariate autoregressive (MAR) model formulates directed interactions (i.e., effective connectivity) between sources. The MAR coefficients (the entries of the MAR matrix) are constrained by the prior knowledge of whole-brain anatomical networks inferred from diffusion MRI. Moreover, to increase the accuracy and robustness of our method, we apply an fMRI prior on the spatial activity patterns and a sparse prior on the MAR coefficients. The observation process of MEG data, the source dynamics, and a series of the priors are combined into a Bayesian framework using a state-space representation. The parameters, such as the source amplitudes and the MAR coefficients, are jointly estimated from a variational Bayesian learning algorithm. By formulating the source dynamics in the context of MEG source r...

Biomedical Engineering / Biomedizinische Technik, 2012

The reconstruction of distributed sources from EEG/MEG data requires additional assumptions on th... more The reconstruction of distributed sources from EEG/MEG data requires additional assumptions on the nature of the underlying sources. While the solution space is often restricted to the cortical sheet, the LORETA method [1,2] additionally imposes covariances between sources based on a smoothness constraint. Functional information, e.g. known from fMRI experiments, can be encoded in the main diagonal elements of a source weighting matrix to bias the solution. Alternatively, priors on the similarity between neighbouring reconstruction points can be used. While it is difficult to provide such a similarity measure for the whole cortex, it is sensible to assume similarity between neighbouring sources and to lift it were a change of functioanatomical properties of the tissue is expected, e.g. at major sulcus lines. We present a new method, LineLORETA, which uses a functio-anatomical inspired source covariance matrix estimated from the combination of similarity information with a general smoothness assumption.

Biomedical Engineering / Biomedizinische Technik, 2012

A System for the Online Reconstruction of Distributed Sources from EEG and MEG data Neuroelectrom... more A System for the Online Reconstruction of Distributed Sources from EEG and MEG data Neuroelectromagnetic signals given by EEG/MEG measurements provide an excellent time resolution of underlying brain processes. A common method to localize brain activity is distributed source reconstruction, particularly in case of analyzing event related potentials. Such analysis requires several preprocessing steps, e.g. artefact detection and correction, filtering, epoch separation and averaging, and is usually performed offline. Online source localization would provide the visualization of brain activity during the measurement, which is interesting for both medical and research applications. We developed a concept that allows to set up and tune an online signal processing chain, including distributed source localization on individual head models. The implemenation is based on OpenWalnut, a software dedicated to multimodal brain visualization [1,2]. The capabilites of online processing were tested based on streaming a previously recorded dataset and measuring the execution times.

2013 IEEE 7th International Conference on Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2013

ABSTRACT Electroencephalography (EEG) and magnetoencephalography (MEG) provide insight into neuro... more ABSTRACT Electroencephalography (EEG) and magnetoencephalography (MEG) provide insight into neuronal processes in the brain in a real-time scale. Brain activity can be modeled in terms of a source distribution found by solving the bioelectromagnetic inverse problem, e.g. using linear source reconstruction methods. Such methods are particularly suitable to be used on modern highly parallel processing systems, such as widely available graphic processing units (GPUs). The utilization of these capabilities paves the way for online neuroelectromagnetic source imaging. We present a system that, according to its modular scheme, can be configured in a very flexible way using graphical building blocks. It allows to use different preprocessing algorithms together with a linear source reconstruction method. The algorithms use both CPU and GPU resources.

International Review of Neurobiology, 2005