Abstract
Attention Deficit/Hyperactivity Disorder (ADHD) is a common mental disorder that exhibits a high incidence rate in children and adolescents, and it is also observed in adults. Currently, there is a lack of objective diagnostic methods for ADHD. Therefore, a three-dimensional residual network (3D-ResNet) deep learning method based on feature extraction from rs-fMRI images for assisting in the diagnosis of ADHD based on resting-state functional magnetic resonance imaging (rs-fMRI) and deep learning models was proposed in this paper. Taking into consideration the temporal characteristics of rs-fMRI, we constructed a 3D-ResNet model based on four-dimensional image. The model utilized TimeDistributed to encapsulate residual blocks which allowed the model to extract spatial features from rs-fMRI while preserving its temporal sequence information. We constructed four different hierarchical structures of 3D-ResNet which are subsequently combined with two different bidirectional recurrent neural networks (BRNNs) to extract sequence features. And BRNNs includes bidirectional long short-term memory (Bi-LSTM) and bidirectional gated recurrent unit (Bi-GRU). The proposed method utilized the ADHD-200 Consortium's public dataset for training and was validated by 5-fold cross-validation. The experimental results indicated that the proposed method in this study demonstrated superior performance on the dataset compared to traditional methods (Accuracy: 76.56%, Sensitivity: 80.16%, Specificity: 90.22%). Therefore, adopting this method can further enhance the accuracy of assisting in the diagnosis of ADHD.