Abstract:Autism spectrum disorder is one of the leading neurodevelopmental disorders in our world, present in over 1% of the population and rapidly increasing in prevalence, yet the condition lacks a robust, objective, and efficient diagnostic. Clinical diagnostic criteria rely on subjective behavioral assessments, which are prone to misdiagnosis as they face limitations in terms of their heterogeneity, specificity, and biases. This study proposes a novel convolutional neural network-based classification tool that aims to identify the potential of different neuroimaging features as autism biomarkers. The model is constructed using a set of sequential layers specifically designed to extract relevant features from brain imaging data. Trained and tested on over 300,000 distinct features across three imaging types, the model shows promise in classifying individuals with autism from typical controls, outperforming metrics of current gold standard diagnostics by achieving an accuracy of 95.4% on a dataset of 1,111 samples with 521 autistic subjects (260 male and 261 female) and 590 controls (297 male and 293 female). 32 optimal features from the training data were identified and classified as candidate biomarkers using an independent samples t-test, in which functional features such as connectivity and the time series of signal intensity from each voxel exhibited the highest mean value differences between individuals with autism and typical control subjects. The p-values of these biomarkers were < 0.001, proving the statistical significance of the results and indicating that this research could pave the way towards the usage of neuroimaging in conjunction with behavioral criteria in clinics.

Identification of Novel Diagnostic Neuroimaging Biomarkers for Autism Spectrum Disorder Through Convolutional Neural Network-Based Analysis of Functional, Structural, and Diffusion Tensor Imaging Data Towards Enhanced Autism Diagnosis