| Popis: |
This paper presents a comprehensive and practical review of autism spectrum disorder (ASD) classification using several traditional machine learning and deep learning methods on data from the Autism Brain Imaging Data Exchange (ABIDE) repository. The objective of this study was to investigate different brain networks and determine their functional connectivity to distinguish between subjects with ASD and those considered typically developing (TD). In the experiments of this paper, functional connectivity was used as a classification feature for 871 resting-state functional magnetic resonance imaging (rs-fMRI) samples collected from the ABIDE repository. The methodology and results of this paper have three main parts.First, we reviewed eight different brain parcellation techniques used for ASD classification from structural, functional, and data-driven perspectives to identify the most promising brain atlas. Second, we evaluate the stability and efficiency of the correlation, partial correlation, and tangent space functional connectivity metrics, and identify the most stable functional connectivity metric. Third, we compared four different supervised learning models used in the ASD classification domain and evaluated the learning performance of each model. In summary, our experimental results show that Bootstrap Analysis of Stable Clusters (BASC) provides the most predictive power for ASD classification, while the correlation metric is the most stable candidate among those models considered. Furthermore, by comparing different classifiers, we conclude that among all the experimentally compared classifiers in this paper, the kernel support vector machine (kSVM) is the optimal classifier for classifying ABIDE fMRI data. The highest sensitivity 64.57% is identified in Table 7. This result was produced using the correlation metric with functional atlas BASC444 and RBF kernel SVM. The corresponding specificity is 73.61%, and the accuracy is 69.43%. Overall, this is the optimal result. |
