Autism is a neurodevelopmental disorder that affects communication and social interactions. Timely identification of autism can lead to earlier and more effective treatments. Given the increasing preva‎lence of autism in recent years, there is a growing need for accurate, fast, and cost-effective diagnostic methods. Traditional methods primarily rely on clinical observations, which are time-consuming and expensive. The use of brain imaging techniques, such as functional magnetic resonance imaging (fMRI), combined with machine learning models, enables the identification of more complex features in the brain's structure and function. This research presents two innovative approaches for autism detection, both based on functional brain data extracted from fMRI signals. These brain signals indicate the blood oxygen level in each brain region, referred to as BOLD signals. In the first approach, wavelet transform was used to extract the frequency levels of BOLD signals. Then, the Tangent-Pearson correlation matrix (TPE) was calculated to extract nonlinear connections between brain regions. Important features were selected using Principal Component Analysis (PCA) and Analysis of Variance (ANOVA), and the Maximum Independence Domain Adaptation (MIDA) algorithm was applied for data harmonization. Finally, a Ridge classifier achieved an accuracy of 71.24% and an AUC of 79.01. In the second approach, the Tangent-Pearson correlation matrix was directly extracted from BOLD signals, and the MIDA algorithm helped reduce dimensions and homogenize the data. To address data imbalance, a Generative Adversarial Network (GAN) with a triple-loss function was used to augment the minority class data. The Kolmogorov-Arnold Network (KAN) was employed to extract features, and a population graph was constructed and fed into the proposed graph network for node classification. This network utilized two separate paths for feature extraction: the first path consisted of Chebyshev blocks, and the second path comprised attention blocks. Chebyshev blocks can capture global structural patterns, while attention blocks are capable of extracting more localized patterns in the population graph. Finally, features extracted from both paths were fused and passed through a KAN network for classification. This method achieved an accuracy of 78.84%, demonstrating its high effectiveness in autism diagnosis.

نادر كريمي

شادرخ سماوي

فرزانه شايق بروجني