fMRIを用いたヒト脳一次聴覚野における1Hz差の深層学習による識別

抄録

Tonotopy plays an important role in sound perception. It functions to activate specific areas of the primary auditory cortex at different frequencies. This study supposes that the brain is activated differently in the primary auditory cortex even for sounds with frequency difference so small that human cannot discriminate them in consciousness. The objective is to use fMRI to capture brain images by sound stimuli with small frequency difference, and to use deep learning to identify the sounds that human hear. In a previous report used the block design and the event-related design for the imaging design. The results showed that the block design produced a high discrimination rate for two sounds with 5 Hz difference. However, the discrimination rate for the two sounds with 1 Hz difference was low for both designs. The purpose of this report is to build a method that can discriminate two sounds with 1 Hz difference, which is low result. As the first step, experiment with increasing the number of training data in the block design. As the second step, it is generally said that the estimation rate improves as the number of training data increases in deep learning. Therefore, as a trial used a combination of both designs in the training phase and each design in the evaluation phase. From the results of a previous study and this study, it was found that the block design is suitable for discriminating two sounds with small frequency difference using deep learning.