To Explore The Clinical Application Value Of Deep Learning Model Based On Conventional Magnetic Resonance Sequence In Identifying Autism Spectrum Disorders

Objective: To explore the clinical feasibility of deep learning model based on cranial conventional MRI in identifying autism spectrum disorders.Methods:We retrospectively analyzed clinical data and conventional magnetic resonance imaging data covering 164 children with autism spectrum disorder（ASD）aged 0-6 and 96 normal children aged 0-6,who were treated in the Affiliated Hospital of Jining Medical College from January 2016 to February 2021.All children with ASD performed the Autism Behavior Scale（ABC）,Childhood Autism Rating Scale（CARS）,and Development Scale for children age 0-6 years [DS（0-6）].First,the deep learning model based on conventional MRI trained in our previous study was used to recognize conventional MRI images of the brain of all subjects and evaluate the diagnostic efficacy of the model.The corresponding diagnostic efficacy was then assessed under four clinical scenarios according to the medical history（positive and negative）and the imaging data（positive and negative）of the study population.Finally,according to the obtained diagnostic results we investigated the differences of the ABC,CARS,and DS（0-6）scores between the ASD identified group and unidentified group.Results: The AUC,accuracy,sensitivity,and specificity of the deep learning model for overall identification of ASD were 0.887,84.62%,82.32%,and 88.54%,respectively.Our subgroup analyses also showed that the deep learning model had high diagnostic efficacy in all four clinical scenarios and its AUC accuracy was highest in the clinical scenario with negative imaging findings and positive medical history,with AUC,accuracy,sensitivity,and specificity of 0.978,94.4%,93.33%,and 100%,respectively.In addition,there were no statistical differences in the scores of the three clinical assessment scales between the ASD identified and unidentified groups.Conclusion: The deep learning model based on the conventional brain MRI achieved good diagnostic efficacy for ASD identification in general and in different clinical scenarios.Moreover,there was no clear correlation between the diagnostic performance of the model and the clinical assessment scale of ASD.