BRAIN DIFFERENCES IN AUTISM SPECTRUM DISORDER

Abstract

Identifying brain abnormalities in autism spectrum disorder (ASD) is critical for early diagnosis and intervention. A novel method Inception 3D-ResNet (iResNet) has been applied to identify brain structural brain differences in ASD with Magnetic Resonance Imaging (MRI) data. However, head motion of MRI has been found to exist and affect the results. Thus, another new method AROMA, which could correct the noise from head motion, has been applied in my research. Background and Context Identifying brain abnormalities in autism spectrum disorder (ASD) is critical for early diagnosis and intervention. To solve this problem, we developed a deep learning-based method, I-ResNet, to identify brain structural abnormalities in ASD. This method can detect structural features from three-dimensional T1-weighted MRI (3D-T1WI) for early diagnosis of ASD. We tested the method on a preschooler ASD (PASD) dataset (n=110) with 75% accuracy in classifying patients from controls. This result was repeated in the Autism Brain Image Data Exchange (ABIDE) dataset (n=1099) with an accuracy of 73%. In combination with the two groups of data, we used I-ResNet method to detect nine regions withsignificant differences between ASD and the control group, including Rolandic operculum, supramarginal gyrus, precentral gyrus, Heschel's gyrus, frontal operculum, superior temporal gyrus, postcentral gyrus, medial orbital superior frontal gyrus, and orbital inferior frontal gyrus. Recent studies have illustrated that head motion-related artifacts remain in resting-state functional Magnetic Resonance Imaging (fMRI) data of ASD even after routine pipeline processing procedures have been applied. However, the effects of head motion data on ASD-fMRI have not been estimated and evaluated by the traditional correction methods. In order to effectively correct head motion of data in ASD, this paper aims to apply Independent Component Analysis (ICA)-based strategy for Automatic Removal of Motion Artifacts (ICA-AROMA) to remove motion artifacts. 306 participants have been enrolled from China and USA. All participants underwent fMRI on a 3D-T1WI. Compared to other traditional methods, we use the ICA-AROMA to denoise for overcoming poor SNR and inconsistency. The removal effects of preprocessed functional images were compared by using four statergies: 6 Head motion parameters (HMP)+2Phys (White matter and cerebrospinal fluid signals), 6HMP+2Phys+ Global Signals Regression (GSR), ICA-AROMA+2Phys and ICA-AROMA+2Phys+GSR. The functional connectivity networks were reconstructed using Thomas Yeo 17 networks parcellations. The effectiveness of each method was measured when the motion-related variance was removed, and the ratio of edges was analyzed. The Quality Control (QC)-functional connectivity (FC) correlation has been found to be statistically significant (p.05, uncorrected and FDR-corrected) in proportion and distance after applying each denoising pipeline. Mean FC was found statistically significant (p.01, uncorrected and FDR-corrected) for Yeo’s 17-Network in each de-noising strategy. The different cortical surfaces have been found to have significant differences between ASD and TD group with the change of different strategy. Compared with other three strategies, the ICA-AROMA head motion corrected FC network was more significant: 22 regions (11 ones: p.05, 11 ones: p.01). Connected with posterior cingulate cortex or postcentral gyrus, more significantly different regions (p.01) have been found between ASD and TD group with other three strategies. Through ICA-Aroma, more meaningful regions can be found , which may help clinicians to identify ASD brain differences from fMRI Data effectively.