Diagnosing autism spectrum disorders using ensemble-aided weighted fused features and attention-based residual LSTM with brain MRI images.

One of the neuro-developmental conditions is called Autism Spectrum Disorder (ASD), which causes changes in the behavior of the patients, and it delays language and social interactions. Details about the functional activity of the brain are provided by Magnetic Resonance Imaging (MRI). Studying each MRI scan of the patients is laborious and time-consuming for doctors and specialists. To tackle these limitations, this paper develops an advanced deep learning diagnosis method. In the beginning, the necessary MRI images are gathered from the available data resource. The input brain images are subjected to an Ensemble Deep Convolutional Neural Network (EDCNN) for feature extraction, which makes the diagnosis easier by reducing the complexities. The ensemble model is created by the integration of the Visual Geometry Group (VGG16), Residual Network (ResNet), and Inception approaches. Further, the resultant features are fused with weights that are optimized using the Improved Random Uniform Number-aided Humboldt Squid Optimization Algorithm (IRUN-HSOA); thus, the weighted fused feature is obtained. The resultant weighted fused feature is fed into Attention-based Residual Long Short-Term Memory (ARLSTM) for the ASD diagnosis. Further, the developed model is compared with different state-of-the-art techniques, and the suitability of the model is discussed for prospects.