Noise-robust Blind Separation of Sources for Optical Imaging of Intrinsic Signals

Schöner, Holger; Stetter, Martin; Schießl, Ingo; Mayhew, J.E.W.; Lund, J.S.; McLoughlin, N.; Obermayer, Klaus

by Holger Schöner, Martin Stetter, Ingo Schießl, J.E.W. Mayhew, J.S. Lund, N. McLoughlin, Klaus Obermayer

Abstract:

Optical imaging of intrinsic signals records small stimulus-related changes in the light reflectance of neuronal tissue, referred to as intrinsic signals, in a strongly noisy environment. The extraction of the two-dimensional neuronal activity pattern from the data requires that the intrinsic signal components which are closely related to neuronal activity (the “mapping signal”) can be reliably separated from spatially global components and from blood vessel artifacts. For this task, we previously proposed a blind source separation (BSS) algorithm (Schiessl et al., Proc. ICA99, pp179-184) which was based on the assumptions that different signal components (i) are separable in space and time, (ii) are smooth functions of space, and (iii) have zero cross-correlation functions. This algorithm used the cross-correlation functions, evaluated at a single shift vector, for the separation of the signal components (sources), and hence was senistive to white noise. Here we propose an improved version for BSS, which finds the optimal separating matrix by iteratively minimizing the whole matrix of cross-correlation functions of the source estimates instead of a single correlation matrix. In contrast to the existing Jacobi-method for BSS, it is not confined to an orthogonal separating matrix, but can find an arbitrary separating matrix that optimally corrects for the influence of the white sensor noise. We demonstrate the superior noise-robustness of our algorithm by comparing its reconstruction error for an artificial data set to that of the Jacobi-method and different single shift methods. We then apply our method to optical imaging records from macaque primary visual cortex (ocular dominance and orientation) and find that it separates the mapping signal very well from global signal components and from vessel artifacts both in differential and in single-condition stacks. Supported by the Wellcome Trust (050080/Z/97)

Reference:

Noise-robust Blind Separation of Sources for Optical Imaging of Intrinsic Signals (Holger Schöner, Martin Stetter, Ingo Schießl, J.E.W. Mayhew, J.S. Lund, N. McLoughlin, Klaus Obermayer), In Society for Neuroscience Abstracts, volume 25, 1999.

Bibtex Entry:

@inproceedings{schoner_noise-robust_1999,
	title = {Noise-robust Blind Separation of Sources for Optical Imaging of Intrinsic Signals},
	volume = {25},
	abstract = {Optical imaging of intrinsic signals records small stimulus-related changes in the light reflectance of neuronal tissue, referred to as intrinsic signals, in a strongly noisy environment. The extraction of the two-dimensional neuronal activity pattern from the data requires that the intrinsic signal components which are closely related to neuronal activity (the “mapping signal”) can be reliably separated from spatially global components and from blood vessel artifacts. For this task, we previously proposed a blind source separation ({BSS}) algorithm (Schiessl et al., Proc. {ICA}99, pp179-184) which was based on the assumptions that different signal components (i) are separable in space and time, (ii) are smooth functions of space, and (iii) have zero cross-correlation functions. This algorithm used the cross-correlation functions, evaluated at a single shift vector, for the separation of the signal components (sources), and hence was senistive to white noise. Here we propose an improved version for {BSS}, which finds the optimal separating matrix by iteratively minimizing the whole matrix of cross-correlation functions of the source estimates instead of a single correlation matrix. In contrast to the existing Jacobi-method for {BSS}, it is not confined to an orthogonal separating matrix, but can find an arbitrary separating matrix that optimally corrects for the influence of the white sensor noise. We demonstrate the superior noise-robustness of our algorithm by comparing its reconstruction error for an artificial data set to that of the Jacobi-method and different single shift methods. We then apply our method to optical imaging records from macaque primary visual cortex (ocular dominance and orientation) and find that it separates the mapping signal very well from global signal components and from vessel artifacts both in differential and in single-condition stacks. Supported by the Wellcome Trust (050080/Z/97)},
	booktitle = {Society for Neuroscience Abstracts},
	author = {Schöner, Holger and Stetter, Martin and Schießl, Ingo and Mayhew, J.E.W. and Lund, J.S. and McLoughlin, N. and Obermayer, Klaus},
	year = {1999},
	keywords = {blind source separation, optical imaging},
	pages = {783}
}