Improvement of MEG source localization performance by Bayesian-based greedy sensor selection
Poster No:
1669
Submission Type:
Abstract Submission
Authors:
Shunsuke Ota1, Kai Miyazaki1, Chihaya Abe2, Keigo Yamada2, Taku Nonomura3, Yoichi Miyawaki1
Institutions:
1The University of Electro-Communications, Chofu, Tokyo, 2Tohoku University, Sendai, Miyagi, 3Nagoya University, Nagoya, Aichi
First Author:
Co-Author(s):
Introduction:
Magnetoencephalography (MEG) can acquire brain activity at a high temporal resolution, but its spatial resolution is not enough to identify the active source location precisely. To resolve this issue, previous studies attempted to develop various source localization methods but it suffers from ill-posedness caused by insufficiency of the number of MEG sensors relative to the potential number of source locations in the brain. However, the increase in the number of sensors is technically difficult in terms of their placement in the limited space, and more importantly, it might not help to improve source localization accuracy because added sensors might provide signals similar to existing sensors. A recent study proposed a method to reconcile this problem by selecting a limited number of sensors with determinant-based greedy-least squares (DG-LS) model [1], showing that MEG signals acquired by full sensors were reconstructed by the small number of selected sensors, though the sensor selection was biased toward near-source locations. Yamada et al. (2021) showed that their Bayesian determinant-based greedy-Bayesian estimation (BDG-BE) can reduce this bias [2]. In this study, we further examined whether the Bayesian-based sensor reduction is also beneficial for achieving better performance in source estimation.
Methods:
We used BDG-BE model to reconstruct MEG signals based on the data from the reduced number of sensors, applied the source estimation model to the reconstructed signals, and compared the results with DG-LS model. Proper orthogonal decomposition was performed to decompose MEG signals into signal (rank: r (< the number of all sensors, n)) and noise components. DG-LS model used the greedy algorithm for selecting sensor locations based on the signal components and then used least-squares estimation with the QR algorithm for MEG signal reconstruction. On the other hand, BDG-BE model considered the noise components and Bayesian prior in addition to the signal components. It selected the sensor locations based on the greedy algorithm and then applied Bayesian estimation for signal reconstruction. In this study, we used two types of Brainstorm tutorial data: the phantom data containing an electrically activated single current dipole out of 32 possible locations, and the real brain activity data obtained while the human subject performed the oddball task. The current dipole was estimated from the reconstructed MEG signals while varying the rank order and the number of sensors using the dipole scanning method provided by Brainstorm software. For the phantom data, source localization accuracy was evaluated by spatial displacement from the true source location for the phantom data and that from the source location estimated using the full MEG signals for the oddball standard stimuli without sensor reduction for the real data at 90 ms after stimulus onset.
Results:
BDG-BE model outperformed DG-LS model in MEG signal reconstruction using a fewer number of sensors. For the phantom data, source localization accuracy was better when using the MEG signals reconstructed from the reduced number of sensors than using the full MEG signals for both models. For the real data, BDG-BE model showed better source localization performance with a fewer number of sensors than DG-LS model when the number of sensors was less than the number of signal modes (Fig. 1, 2).
Conclusions:
We demonstrated that MEG signals can be reconstructed even after the number of sensors is reduced, and the sensor reduction might be even useful to improve the source localization accuracy. Source localization with sensor reduction further showed that BDG-BE model outperformed DG-LS model for the real data that might have complex signal components, which probably reflected in the number of signal modes, than the simple phantom data. These results suggest that the proper MEG sensor selection can improve source localization accuracy and is particularly useful for analyzing complex brain activity patterns.
Modeling and Analysis Methods:
Bayesian Modeling
EEG/MEG Modeling and Analysis 1
Methods Development 2
Novel Imaging Acquisition Methods:
MEG
Keywords:
Computational Neuroscience
MEG
Modeling
Source Localization
1|2Indicates the priority used for review
Provide references using author date format
[2] Yamada, K., Saito, Y., Nankai, K., Nonomura, T., Asai, K. and Tsubakino, D. (2021) ‘Fast greedy optimization of sensor selection in measurement with correlated noise’, Mechanical Systems and Signal Processing, vol. 158, article 107619.