Automated Autism Spectrum Disorder Diagnosis using Graph Metrics from Diffusion Tensor Imaging and Machine Learning.

Researchers used brain scans and computer analysis to try to improve autism diagnosis. They looked at how different parts of the brain connect to each other in autistic people compared to non-autistic people. Their computer program could correctly identify autism in about 8 out of 10 cases by analyzing these brain connections. This could help doctors diagnose autism more accurately in the future, though more research is needed.

This study investigated whether brain imaging combined with machine learning could improve autism diagnosis. Researchers analyzed diffusion tensor imaging (DTI) data from the ABIDE-II database, examining white matter connections in the brain. They created mathematical models of brain networks and extracted 300 features representing how different brain regions connect. Using machine learning algorithms (logistic regression and support vector machines), they achieved 82.34% accuracy in distinguishing autistic individuals from typically developing participants.

Key distinguishing features included specific white matter regions: the left cingulum, left anterior corona radiata, and corpus callosum. The approach aims to develop objective diagnostic tools to supplement current behavioral assessments for autism spectrum disorder diagnosis.

Shows promise for developing objective autism diagnostic tools using brain imaging and artificial intelligence. However, requires validation in larger, diverse populations and comparison with standard diagnostic methods before clinical implementation. Current accuracy levels may supplement but not replace comprehensive clinical assessment.

Sample size not reported. Single dataset used (ABIDE-II). No validation on independent populations. Conference paper format suggests preliminary findings. Unclear participant demographics and diagnostic criteria. No comparison to clinical diagnostic accuracy rates.

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition with an increasing global prevalence, yet its diagnosis remains challenging due to the absence of objective biomarkers and reliance on subjective behavioral assessments. This study aims to bridge this gap by integrating advanced neuroimaging techniques, graph theory, and machine learning algorithms to develop a diagnostic classification model for ASD. Initially, diffusion tensor imaging (DTI) data from individuals with ASD and typically developing (TD) participants were obtained from the Autism Brain Imaging Data Exchange-II (ABIDE-II) database. The data were preprocessed, followed by the extraction of DTI-derived parameters from various white matter regions of the brain.

A structural correlation matrix was constructed using a Pearson correlation method. Further, graphs were generated from the matrix to model brain organization by representing regions as nodes and their structural correlations as edges. We computed six graph metrics, including betweenness centrality, closeness centrality, clustering coefficient, degree centrality, participation coefficient, and strength from the graph network, which leads to a total of 300 features per individual. Finally, we built the diagnostic classification models using logistic regression and support vector machines (SVM) and the performance of the models were evaluated.

Our results revealed that SVM produced the highest classification accuracy of 82.34% with 225 graph-theoretical features. The top three distinguishing features for ASD classification were strength of the cingulum left, closeness centrality of the anterior corona radiata left, and betweenness centrality of the genu of the corpus callosum. Our approach provides insights into ASD-related alterations in brain structural networks and contributes toward the development of objective diagnostic tools.Clinical Relevance-This study highlights the potential of DTI-based graph-theoretical metrics combined with machine learning classifiers to differentiate ASD from TD participants.