Advancing ASD detection: novel approach integrating attention graph neural networks and crossover boosted meerkat optimization

Autism spectrum disorder (ASD) is a neurodevelopmental condition that significantly impacts the lives of many children due to its hidden symptoms. Early detection of ASD is challenging because of its complex and heterogeneous nature. Magnetic resonance imaging (MRI) has emerged as a crucial tool for early detection, offering non-invasive imaging with detailed soft tissue information. However, existing approaches face limitations such as overfitting, underfitting, class imbalance, control, domain shift, and behavioral issues. To address these challenges, this paper proposes a novel ASD detection and classification model called the Autism Spectrum Disorder-based Attention Graph Neural Network and Crossover Boosted Meerkat Optimization (ASD-AttGCBMO) algorithm. The proposed method utilizes structural Magnetic Resonance Imaging (sMRI) data from the ABIDE 1 dataset. The data undergoes preprocessing to remove artifacts and noise, ensuring high image quality and consistency. Node feature extraction employs voxel-based morphometry (VBM) and surface-based analysis, which extract relevant features such as surface area, cortical thickness, shape descriptors, and brain volumes. The ASD-AttGCBMO model is trained using preprocessed sMRI images, employing the Adam and Stochastic Gradient Descent (SGD) optimizers to prevent overfitting, reduce classification loss, and improve convergence. The model is designed to enhance the learning process and capture complex patterns for accurate feature classification between ASD and control subjects. To optimize the hyperparameters in the attention-based neural network model, the CBMO algorithm is employed. Experimental validation is conducted using essential performance evaluation measures. The proposed method achieves impressive results, with accuracy, precision, recall, specificity, F1-score, Area under Receiver Operating curve (AUC/ROC), and computational time values of 98.8%, 99%, 98.5%, 98.6%, 98.2%, 0.989, and 3.05 s, respectively. Comparative analysis demonstrates that the proposed method outperforms other state-of-the-art methods.