Artificial intelligence networks combining histopathology and machine learning can extract axon pathology in autism spectrum disorder
bioRxiv – October 25, 2024
Source: medRxiv/bioRxiv/arXiv
Summary
Research reveals that axon structures in the brain differ significantly between neurotypical individuals and those with autism. By harnessing machine learning, scientists developed a method to analyze brain tissue more efficiently, achieving 98% accuracy in classifying white matter regions. This innovative approach promises to enhance our understanding of autism-related brain changes.
Abstract
Axon features that underlie the structural and functional organization of cortical pathways have distinct patterns in the brains of neurotypical controls (CTR) compared to individuals with Autism Spectrum Disorder (ASD). However, detailed axon study demands labor-intensive surveys and time-consuming analysis of microscopic sections from post-mortem human brain tissue, making it challenging to systematically examine large regions of the brain. To address these challenges, we developed an approach that uses machine learning to automatically classify microscopic sections from ASD and CTR brains, while also considering different white matter regions: superficial white matter (SWM), which contains a majority of axons that connect nearby cortical areas, and deep white matter (DWM), which is comprised exclusively by axons that participate in long-range pathways. The result was a deep neural network that can successfully classify the white matter below the anterior cingulate cortex (ACC) of ASD and CTR groups with 98% accuracy, while also distinguishing between DWM and SWM pathway composition with high average accuracy, up to 80%. Multidimensional scaling analysis and sensitivity maps further underscored the reliability of ASD vs CTR classification, based on the consistency of axon pathology, while highlighting the important role of white matter location that constrains pathway dysfunction, based on several shared anatomical markers. Large datasets that can be used to expand training, validation, and testing of this network have the potential to automate high-resolution microscopic analysis of post-mortem brain tissue, so that it can be used to systematically study white matter across brain regions in health and disease.