| Summary: | 碩士 === 國立交通大學 === 生醫工程研究所 === 107 === Magnetoencephalography (MEG) is a noninvasive functional brain imaging tool that measures the magnetic field produced by brain activities. With superior temporal resolution, functional brain research in both clinical and basic neuroscience fields can be promoted. Decoding the message in the mind from brain signals is always a interesting research topic. In the past, brain scientists tried many methods to present the brain activity by music. They extracted the features of brain signals with traditional methods, such as short-time Fourier transform, mean, variance, and wave amplitude, and defined their own mapping rules to transform these features into music in MIDI format.
In our experiment, we used the MEG signals from Taipei Veterans General Hospital. Data of eight percussionists were used in this study. In each trial, the participants were asked to follow the instruction displayed on the screen, and we only take the listening part only for our analysis. In this work, we want to reconstruct the exact rhythm we heard when the MEG signals were collected. We proposed a system based on neural networks that consist of a classifier and a generator. With the help of the accurate classifier, the following generator can reconstruction the rhythm from the features extracted from MEG signals.
According to our experimental results, the trained classifier can accurately classify the MEG signals into different rhythm stimuli. The average classification accuracy is 70.83%. Moreover, it can also extract the features from MEG signals that can reconstruct the rhythm using the generator. The average reconstruction accuracy is 52.78%. However, if we only consider those that can be accurately classified, the average accuracy is 69.47%.
In summary, the proposed classifier can accurately classify the MEG signals and extract the useful features for reconstruction. With the features, the proposed generator can reconstruct the rhythm in waveform directly with certain accuracy.
|
|---|
