Transformer-based Model for fMRI Data: ABIDE Results

Functional magnetic resonance imaging (fMRI), which is a non-invasive technique for measuring brain signals, is widely used in the diagnosis of Alzheimer's disease (AD), Autistic spectrum disorder (ASD), Depression, and other brain neurological diseases. In the current research, fMRI signals are usually analyzed using convolutional neural network (CNN) and recurrent neural network (RNN). However, these models are ineffective when considering the relationship between nonadjacent brain regions and analyzing long-time span fMRI signals. Transformer, which completely abandons the architecture of the traditional neural network, can overcome the above limitations through positional decoding and self-attention mechanism. In this study, a transformer-based model is proposed, which is the first time to apply transformer to fMRI data analysis. In addition, positional decoding is an essential part of transformer. Functional connection matrix is creatively used as the positional decoding of transformer-based model for fMRI data. The autism brain imaging data exchange (ABIDE) dataset is used to evaluate this model. The transformer-based model achieved a classification accuracy of 74.18% using subject-wise 10-fold cross-validation. It cannot exceed the classification accuracy (79.50%) of the state-of-the-art model based on hand-engineered features but exceed the highest classification accuracy (74.00%) of the model based on original fMRI image data in ABIDE dataset. This study provides a transformer-based model for original fMRI data, which is helpful to realize the early diagnosis of ASD, AD, Depression, and other neurological diseases.