Decoding Music Genres Based on High Resolution Brain Activity Information

Decoding stimuli from the brain is important for understanding the brain's processing mechanism of external information, which can also promote the development of brain-computer interface. Most of the existing researches focused on the decoding of audiovisual information. Few studies investigated the decoding of music stimuli, and the decoding accuracy is also not satisfactory. This paper uses a public 7-Tesla fMRI image dataset, which collects the high resolution blood oxygen dependent level (BOLD) signals when 20 subjects listen to 5 music genres. After fMRI data preprocessing, two feature selection methods were used. One is based on a prior template (called MASK) including the Heschl's gyrus (HG), the anterior superior temporal gyrus (aSTG), and the posterior superior temporal gyrus (pSTG). The other one is whole brain analysis of variance (ANOVA). Then, the gradient boosting decision tree (GBDT) algorithm is used to train the decoding models to discriminate different music genres. Results showed that among the five genres, ambient music is easier than the other four categories to be decoded. Compared with the previous study that used the same dataset and the same prior template but combined with the classifier of support vector machine[1], the GBDT algorithm improved the accuracies in most genres from around 45% to around 65%. Compared with the MASK method, the ANOVA method improved the decoding accuracy to larger than 74% for all genres. Analysis of contributive regions in the ANOVA method shows that insular and parietal regions are additionally recruited in decoding of music genres, which may be related to music understanding and emotion expression. In summary, the findings in this paper help to understand the brain mechanism of music processing in depth. Meanwhile, the decoding model proposed in this study can also be used to classify other fMRI stimuli.