Advancing ASD Diagnostic Classification with Features of Continuous Wavelet Transform of fMRIand Machine Learning Algorithms

Abstract

In this study, we aimed to develop a diagnostic classification model for autism spectrum disorder (ASD) using scalograms of functional magnetic resonance imaging (fMRI) data. Initially, the fMRI data from the ABIDE-I and ABIDE-II databases were preprocessed, and brain regions of interest (ROIs) were parcellated. Subsequently, scalograms were generated using Continuous Wavelet Transform from the average time series of each ROI. Various features were extracted from these scalogram images, including gray-level co-occurrence matrix, gray-level run-length matrix, fractal dimension texture analysis, Zernike's moments, Hu's moments, and first-order statistics. We employed CatBoost and Decision Tree machine learning classifiers and performed extensive five-fold cross-validation with hyperparameter tuning to evaluate the model's performance. The results indicated an overall accuracy of 83.04%, sensitivity of 66.66%, specificity of 85.57%, and an F1-score of 75.85% for the Decision Tree classifier, particularly a region in the Retro-Splenial Temporal network. In addition, regions from networks such as fronto-parietal task control network, Ventral Attention, and Visual, showcased promising classification performance for ASD. These findings highlight the efficacy of the proposed model in providing accurate ASD classification, which holds significant implications for early diagnosis and intervention strategies. © 2023 IEEE.