Decision Level Fusion for Diagnosing Autism Spectrum Disorder

Abstract

Automated diagnosis of Autism Spectrum Disorder(ASD) by integrating Machine Learning (ML) techniques is rapidly growing in the field of neuroscience. In this study, we proposed an unsupervised approach for diagnosing ASD with Deep Learning (DL) models such as UNet, GAN, and SAGAN. The axial and coronal slices of T1-weighted longitudinal Structural Magnetic Resonance Imaging (sMRI) from multisite ABIDE II are used for the study. At first, the DL models are trained only with Typical Development (TD) subjects to reconstruct multiple slices, and then we used both ASD and TD subjects for testing. outliers are detected using a combination of L2 loss and cosine similarity loss. Finally, individual classification results from axial and coronal slices are fused at the decision level using maximum probability yielding classification accuracy of 95.65% and an AUC score of 0.90.