Deep Atlas Fusion for Improved Autism Spectrum Disorder Detection

Autism spectrum disorder (ASD) influences the brain and nervous system, manifesting in diverse ways and varying in severity among individuals. Some individuals with ASD may encounter considerable difficulties that necessitate comprehensive support, while others may display less pronounced symptoms and possess the ability to function independently. The majority of current methods use resting-state functional magnetic resonance imaging (rs-fMRI) to diagnose ASD. This neuroimaging technique examines brain activity and allows researchers to examine functional connectivity (FC) patterns among distinct brain areas. Most existing methods for analyzing the brains of individuals with ASD extract features from specific regions of interest (ROIs) defined in various brain atlases. These atlases enable researchers to compare and quantify anatomical and functional variations in the brains of different individuals. Many current approaches face limitations because they rely on a single atlas, which may not adequately uncover intricate brain connectivities and functional variations among distinct brain regions. This can result in suboptimal diagnostic performance. To overcome these challenges, this paper employs multiple atlases to partition the brain into various ROIs. Additionally, a deep neural network (DNN) is used, which is based on a deep fusion between the outputs of four different atlases. This fusion method enables the updating of the model's parameters using information from all atlases, making it a trainable ensemble classifier. The multi-atlas fusion technique described in this paper has a binary ASD detection accuracy of 76.78%, which outperforms the state-of-the-art results on the same autism brain imaging data exchange I (ABIDE I) dataset.