Abstract
•A new ADHD classification method based on the spatio-temporal features of fMRI was proposed.•The nested residual convolutional denoising autoencoder have more power to represent the spatial features.•The 3D convolutional gated recurrent unit is first employed to extract the spatio-temporal features.
Attention deficit/hyperactivity disorder is a neurological disorder characterized by inattention, hyperactivity and impulsivity. Since the resting functional magnetic resonance imaging (rs-fMRI) can reflect the brain activity state, automatic diagnosis of attention deficit/hyperactivity disorder (ADHD) based on re-fMRI has become a critical method which can relieve the pressure of doctors meanwhile improve the efficiency and accuracy of diagnosis. Up to now, the traditional diagnosis of ADHD based on rs-fMRI converts rs-fMRI into functional connectivity or low-dimensional data for classification which will lead to original information loss since rs-fMRI contains 3D spatial and 1D temporal information. In this paper, we propose a deep learning model consisting of the spatio-temporal feature extraction and classifier. The nested residual convolutional denoising autoencoder (NRCDAE) is exploited to reduce the spatial dimension of rs-fMRI and extract the 3D spatial features. Then, the 3D convolutional gated recurrent unit (GRU) is adopted to extract the spatial and temporal features simultaneously. The spatio-temporal features extracted are sent into the sigmoid classifier for classification. The training and testing are carried out across five sites of ADHD-200 and the results of experimental show that the proposed method can effectively improve the classification performance between ADHD and typically developing (TD) in the cross-site test set from 1.14% to 10.9% compared with other methods. To our knowledge, the 3D convolutional GRU is first employed to extract the features of the fMRI and the model has better generalization than other models.