Source Localization Applying H(Div) Source Model and Hierarchical Bayes Model in Eeg and Meg Combining Machine Learning
He, Qin (2019)
2019
Materiaalitekniikka
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2019-05-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201901251162
https://urn.fi/URN:NBN:fi:tty-201901251162
Tiivistelmä
In this work, the focus is on the reconstruction of the dipole source on electro-/magnetoencephalography (E/MEG) which is conducted utilizing a MATLAB software Zeffiro. We applied the finite element method (FEM) in the forward model and a hierarchical Bayes model in solving the inverse problem. To be more specific, the iterated max prior algorithm (IAS MAP) is utilized. After the methodology of the models, the introduction and main functions of the Zeffiro is covered as well. With adjusting the different parameters, including low (0.02) and high (0.05) noise level, gamma hyperprior and inverse gamma hyperprior, small (10−9) and large (10−5)
scaling parameters, we discovered different accuracies of the dipole source reconstruction on EEG and MEG data. To summarize, the better location accuracy of the dipole is usually obtained at the expense of the less accuracy of the angle. The magnitude reconstruction is not well reconstructed except that in the deep source (thalamus) in EEG signal. Especially, the reconstruction in terms of magnitude for MEG basically all failed. (The reconstruction of the amplitude is averagely 100 times of the original amplitude). In general, the reconstruction in deep source is better in EEG signal while that in superficial (somatosensory) source is better in MEG. In addition, the gamma hyperprior works for deep source while inverse gamma works better for superficial source in location and reverse in angle, except for the superficial single dipole in MEG data. For two sources, the results of the reconstruction were inferior in angle accuracy. Basically, local and global effects can be distinshed by position differences and angle differences respectively. Because the position differences can better be detected by electrodes (coils) closer to the dipole given the density of the magnetic field is higher while the angle differences are larger for the electrodes (coils) further to the dipole given the angle is radiatively larger along the same origin direction. Thus, one of the hyperprior seems to work better for the local and global effects. Data analysis skills is utilized in the visualization of the results. The grouped scatters show the relation between dipole source, location/ best scaling parameter and amplitude difference, location difference and angle difference, the outliers are detected through SVM classifier. Finally, the dipole location is predicted as the reference of the inverse problem. In the future, machine learning can also be utilized to evaluate the inverse model.
scaling parameters, we discovered different accuracies of the dipole source reconstruction on EEG and MEG data. To summarize, the better location accuracy of the dipole is usually obtained at the expense of the less accuracy of the angle. The magnitude reconstruction is not well reconstructed except that in the deep source (thalamus) in EEG signal. Especially, the reconstruction in terms of magnitude for MEG basically all failed. (The reconstruction of the amplitude is averagely 100 times of the original amplitude). In general, the reconstruction in deep source is better in EEG signal while that in superficial (somatosensory) source is better in MEG. In addition, the gamma hyperprior works for deep source while inverse gamma works better for superficial source in location and reverse in angle, except for the superficial single dipole in MEG data. For two sources, the results of the reconstruction were inferior in angle accuracy. Basically, local and global effects can be distinshed by position differences and angle differences respectively. Because the position differences can better be detected by electrodes (coils) closer to the dipole given the density of the magnetic field is higher while the angle differences are larger for the electrodes (coils) further to the dipole given the angle is radiatively larger along the same origin direction. Thus, one of the hyperprior seems to work better for the local and global effects. Data analysis skills is utilized in the visualization of the results. The grouped scatters show the relation between dipole source, location/ best scaling parameter and amplitude difference, location difference and angle difference, the outliers are detected through SVM classifier. Finally, the dipole location is predicted as the reference of the inverse problem. In the future, machine learning can also be utilized to evaluate the inverse model.