Precision Diagnosis of Autism Through Resting FMRI and Machine Learning

In this paper, we present a pioneering diagnostic system that utilizes resting state Functional Magnetic Resonance Imaging (fMRI) and Machine Learning (ML) to effectively evaluate individuals with autism spectrum disorder (ASD). Our system’s primary objective is to assess the severity of autism by accurately identifying specific brain regions that demonstrate robust correlations with the behavioral patterns observed in autistic subjects. The proposed system involves several essential steps: i) It begins by preprocessing fMRI data to correct head motion, reduce susceptibility distortion in the reconstructed BOLD signal, and align the subject’s fMRI with their structural MRI, thereby improving fMRI resolution. ii) Following this, the brain cortex is extracted from the aligned fMRI using its sMRI. iii) The brain is then divided into 76 areas per hemisphere using MNI152 standard space. To identify brain areas associated with ASD, fMRI radiomics employs estimating the correlation matrix, outlining the mutual synchronization between each pair of brain regions. v) Areas significantly correlated with ASD, at a 95% confidence interval, are pinpointed using the Recursive Feature Elimination (RFE) algorithm. vi) Lastly, the final step involves using the Linear support vector machine (LSVM) to diagnose each subject as normal or autistic, determining their respective severity levels, and identifying abnormal brain regions crucial in detecting abnormal neurocircuits that play a pivotal role in managing autism. The proposed approach was tested on 344 with ASD and 374 typically developing individuals from the Autism Brain Imaging Data Exchange II (ABIDE II). Through 5-fold cross-validation, the proposed system attained 97% accuracy, 90% sensitivity and 0.99% specificity.