GACT: A Two-Stage Age Prediction Model Combining a Global Attention Block

The use of neuroimaging data to estimate brain age plays a critical role in understanding normative brain development and the progression of neurological disorders. While deep learning models based on fMRI data have achieved success in this field, most studies rely on spatial maps or functional connectivity as input features. Despite significant progress, these methods often lose fine-grained information about the brain. Therefore, our study proposes a novel approach that directly utilizes unsegmented fMRI data as input features to better leverage the spatiotemporal information of fMRI data. Our method integrates convolutional neural networks (CNN) and transformer models to capture spatial and temporal features simultaneously. At the final stage, the extracted spatiotemporal features are fed into a Multi-Layer Perceptron (MLP) for age prediction. Experimental results demonstrate the outstanding performance of our model in age prediction tasks and identify regions significantly impacting age regression tasks through an explainability method, offering new insights and methodologies for future neuroimaging research.