Diagnosis of Autism Spectrum Disorder based on Denoised Multiple Age-Specific Structural Features
Poster No:
366
Submission Type:
Abstract Submission
Authors:
Dongyue Zhou1, Yunge Zhang1, Wei Zhao1, Yuxing Hao1,2, Fengyu Cong1, Huanjie Li1
Institutions:
1School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology, Dalian, Liaoning, 2University of Jyväskylä, Jyväskylä, Finland
First Author:
Dongyue Zhou
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
Co-Author(s):
Yunge Zhang
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
Wei Zhao
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
Yuxing Hao
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology|University of Jyväskylä
Dalian, Liaoning|Jyväskylä, Finland
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology|University of Jyväskylä
Dalian, Liaoning|Jyväskylä, Finland
Fengyu Cong
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
Huanjie Li
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology
Dalian, Liaoning
Introduction:
Diagnosis of Autism Spectrum Disorder (ASD) based on sMRI is more objective than clinical scales due to the high heterogeneity in symptom severity. Existing studies in binary classification between ASD and control based on small sample reported remarkable classification accuracies. However, accuracies on large heterogeneous datasets were not high[1,2]. This may be due to insufficiently specific feature selection, insufficiently effective feature combination, and the presence of multi-site noise. In this paper, we applied DP-ICA method to remove the site effect, and then combined three kinds of age-specific structural features including regional, interregional features and disease vulnerability index using multiple kernel learning (MKL) for binary classification task. Results showed that our procedure reached the accuracy of 85.63% when discriminating ASD children from control.
Methods:
Structural MRI data of 660 subjects from public ABIDE II dataset[3] were included in this study. We divided all people into three age groups (Age1: 340 children aged 6-12 years old; Age2: 170 adolescent aged 12-18 years old; Age3: 150 adults over the age of 18), and ensured that ASD and control group in each age stage had the same number of people and matched age (no difference between two groups).
Data preprocessing was conducted by FreeSurfer. Desikan-Killiany Atlas[4] was used here for extracting three types of characteristics. (1) Morphological features (MF) included cortical thickness (CT), pial surface area (PSA), grey matter volume (GMV), folding index (FI) and curvature (CURV) were calculated in native space. (2) We constructed individual morphological brain network (MBN)[5] with each subject's grey matter maps processed by FSL-VBM. Next, DP-ICA method[6] was used for correcting site effects on MF and MBN. (3) The third type of features - regional vulnerability index (RVI)[7], which was quantified by the Pearson correlation between standardized individual brain values and the effect size from large sample meta-analysis. We performed the meta-analysis by calculating the effect size in each site separately (16 sites in our study), and then combing them (refer to [8]).
After preparing these 3 features, we performed a two-step feature selection separately in 3 age stages for obtaining optimal feature subsets. Firstly, we applied two sample t-test on two high-dimensional features (MF and MBN), and features with p>0.05 (uncorrected) were excluded. Then, SVM-based recursive feature elimination (SVM-RFE) was conducted to evaluate the importance of these features in classification. Finally, we utilized MKL[9] with appropriate weight to combine selected and specific MF, MBN and RVI in 3 age stages. We calculated accuracy, sensitivity, specificity, area under receiver operating characteristic curve and F-score for assessment of classification performance in a five-fold cross validation. And we repeated this procedure for 100 times to evaluate the performance of MKL compared to others via a paired t-test.
Data preprocessing was conducted by FreeSurfer. Desikan-Killiany Atlas[4] was used here for extracting three types of characteristics. (1) Morphological features (MF) included cortical thickness (CT), pial surface area (PSA), grey matter volume (GMV), folding index (FI) and curvature (CURV) were calculated in native space. (2) We constructed individual morphological brain network (MBN)[5] with each subject's grey matter maps processed by FSL-VBM. Next, DP-ICA method[6] was used for correcting site effects on MF and MBN. (3) The third type of features - regional vulnerability index (RVI)[7], which was quantified by the Pearson correlation between standardized individual brain values and the effect size from large sample meta-analysis. We performed the meta-analysis by calculating the effect size in each site separately (16 sites in our study), and then combing them (refer to [8]).
After preparing these 3 features, we performed a two-step feature selection separately in 3 age stages for obtaining optimal feature subsets. Firstly, we applied two sample t-test on two high-dimensional features (MF and MBN), and features with p>0.05 (uncorrected) were excluded. Then, SVM-based recursive feature elimination (SVM-RFE) was conducted to evaluate the importance of these features in classification. Finally, we utilized MKL[9] with appropriate weight to combine selected and specific MF, MBN and RVI in 3 age stages. We calculated accuracy, sensitivity, specificity, area under receiver operating characteristic curve and F-score for assessment of classification performance in a five-fold cross validation. And we repeated this procedure for 100 times to evaluate the performance of MKL compared to others via a paired t-test.
Results:
Fig.1 showed the superiority of MKL in combination of 3 different kinds of features, and the advantage of age-specific features in classification. It's obvious that age1-specific features performed better than that in all subjects. And RVI had the highest weight in each age stage.
Fig.2 displayed the most discriminative and age-specific features in 3 age stages. We could observe the same and specific characteristics among them.
Fig.2 displayed the most discriminative and age-specific features in 3 age stages. We could observe the same and specific characteristics among them.
·Fig.1 Classification performance based on multi-kernel learning (MKL).
·Fig.2 The most discriminative features in three age stages.
Conclusions:
We proposed a procedure that combined three kinds of age-specific structural features in each age stage using MKL to distinguish ASD and control. Results showed that the most identifiable features differed in three age stages, thus the method sub-grouped by age was effective, especially in children group, where the accuracy was highest, indicating that abnormalities in ASD may be more easily observed in childhood.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural)
Novel Imaging Acquisition Methods:
Anatomical MRI 2
Keywords:
Autism
Machine Learning
STRUCTURAL MRI
Other - Multi-kernel learning
1|2Indicates the priority used for review
Provide references using author date format
2. Katuwal, Gajendra J. (2016), 'Divide and Conquer: Sub-Grouping of ASD Improves ASD Detection Based on Brain Morphometry.', PLOS ONE, 11 (4): e0153331.
3. Di Martino, A. (2014), 'The Autism Brain Imaging Data Exchange: Towards a Large-Scale Evaluation of the Intrinsic Brain Architecture in Autism.', Molecular Psychiatry, 19 (6): 659–67.
4. Desikan, Rahul S. (2006), 'An Automated Labeling System for Subdividing the Human Cerebral Cortex on MRI Scans into Gyral Based Regions of Interest.', NeuroImage, 31 (3): 968–80.
5. He, Changchun. (2021), 'Individual‐based Morphological Brain Network Organization and Its Association with Autistic Symptoms in Young Children with Autism Spectrum Disorder.', Human Brain Mapping, 42 (10): 3282–94.
6. Hao, Yuxing. (2023), 'Site Effects Depth Denoising and Signal Enhancement Using Dual-Projection Based ICA Model.' Preprint. Neuroscience.
7. Kochunov, Peter, L. (2022), 'ENIGMA‐DTI: Translating Reproducible White Matter Deficits into Personalized Vulnerability Metrics in Cross‐diagnostic Psychiatric Research.', Human Brain Mapping, 43 (1): 194–206.
8. Xie, Yapei. (2022), 'Alterations in Connectome Dynamics in Autism Spectrum Disorder: A Harmonized Mega- and Meta-Analysis Study Using the Autism Brain Imaging Data Exchange Dataset.', Biological Psychiatry, 91 (11): 945–55.
9. Wee, Chong-Yaw. (2012), 'Identification of MCI Individuals Using Structural and Functional Connectivity Networks.', NeuroImage, 59 (3): 2045–56.