Detecting ASD in individuals based on their brain functional connectivity patterns

Abstract

The main focus of this research is on accurate pediatric neurodevelopmental disorder diagnostics as well as early detection of Autism Spectrum Disorder (ASD) children and adolescents. Given the heightened neuroplasticity of the brain during these formative years, the ability to detect ASD (Autism Spectrum Disorder) early in development has profound implications for intervention and outcomes.This inherent variability poses a challenge when attempting to classify ASD (Autism Spectrum Disorder) using the intricate 4D fMRI data from ABIDE I that captures both spatial and temporal aspects of brain activity. Majority of research dealing with the neuroimaging methods has typically reduced this high-dimensional 4D data to a 2D correlation matrix, leading to a potential loss of crucial information related to both the dynamic brain activity and the connectivity patterns. In response to this limitation, we introduce an innovative preprocessing framework, NeuroSeg3D (Neuroimaging Segmentation for 3D framework), a cutting-edge pipeline designed to preserve the full 4D integrity of the fMRI(Voxel intensity) data. Also by retaining the inherent temporal and spatial complexities of the 4D fMRI scans, NeuroSeg3D empowers the model to extract a more nuanced understanding of the brain’s architecture, which is paramount for accurate ASD classification. Building on this advanced preprocessing methodology, we developed CASD-Net (Children ASD Neural Network), a specialized convolutional neural network (CNN) meticulously engineered to handle the complexities of 4D neuroimaging data. CASD-Net (Children ASD Neural Network) utilizes state-of-the-art deep learning techniques to extract high-level features from raw fMRI volumes, capturing both spatial and temporal dependencies crucial for distinguishing ASD from non-ASD individuals. By employing multiple convolutional layers, attention mechanisms, and optimization strategies, CASD-Net(Children ASD Neural Network) is able to identify and learn intricate patterns within the 4D brain scans that are often imperceptible to conventional analysis methods. Beside this the model’s performance was rigorously evaluated through both test set validation and 5-fold cross-validation, ensuring its robustness and generalizability across different subsets of data. Most importantly CASD-Net(Children ASD Neural Network) has achieved a remarkable test accuracy of 93.97%, underscoring its high performance in distinguishing ASD from control subjects. Beyond that, the cross-validation results of the model depicted a solid generalization capabilities, with an average accuracy of 83.57%, an average loss of 0.3147, and a standard deviation of 16.27%, and thus highlighting the model’s adaptability despite variability in the dataset. The model we developed has demonstrated a robust performance through the classification metrics like precision, recall, and F1-score. Ensuring that both ASD(Autism Spectrum Disorder) and control classifications were accurate, reliable, and balanced. This result has underscored an enormous potentiality that our developed model possessed while revolutionizing neuroimaging diagnostics, especially with the complex and high-dimensional data ABIDE I.