Autism Spectrum Disorder Diagnosis Using Neuroimaging Based on Convolutional Neural Networks and Particle Swarm Optimisation

This study presents an automated diagnosis system using structural magnetic resonance images (sMRI) for autism spectrum disorder (ASD). The system is built around a Con-volutional neural network (CNN) architecture, this framework optimises its hyperparameters for improved performance using Particle swarm optimisation (PSO). Preprocessing and training are the two main stages of the model's development. During the preprocessing stage, black and hazy images were removed from the dataset to maintain data integrity. The remaining images were then scaled to fit the model's required input dimensions. In order to maximise learning and maintain the differentiation between normally developing (TD) and ASD groups, it was imperative to take this step to guarantee that the data input into the CNN was uniform and clean. The learning rate and network depth of the CNN are two hyperparameters that are adjusted during the training phase using PSO, and it finds the best configurations by imitating the behaviour of swarms. As a result, the suggested diagnosis approach for Autism Spectrum Disorder that makes use of structural MRI and a CNN that has been upgraded using PSO achieves an outstanding 85.0% accuracy rate. The reliability of the model is configured through five-fold cross-validation, ensuring its robustness and consistency. The model's performance is further highlighted by comparing it with recent studies and commonly used optimizing models, where it demonstrates significant improvements in accuracy and diagnostic efficiency.