A Novel Generalizable Neural Decoder for Visual Object Recognition Using fMRI Data of Human Brains

Understanding the activity and functioning of various brain regions is a crucial aspect of neuroscience. Functional Magnetic Resonance Imaging (fMRI) has emerged as a powerful tool for recording brain activity during specific tasks, providing valuable insights into brain function. With advancements in machine learning and statistics, there is a growing potential to process and identify patterns in fMRI data. In this study, we aimed to decode the information embedded in fMRI data to recognize objects presented to subjects as visual images. We proposed a novel solution using deep learning techniques to develop a generalized model to address the limitations of previous subject-specific models. We introduced a specially designed transformer model for fMRI data, which not only exhibited superior spatial pattern recognition capabilities but also considered the relationships among different Regions Of Interests (ROIs) involved in visual processing. This approach ensured adaptability to differences in brain structure across different subjects. Our results demonstrated the potential for decoding visual images based on the brain activity of new subjects, thereby opening up a new way in reading the brain. The proposed decoder can be used in brain machine interface systems to help people with visual disabilities