Optimized quadrangle attention consolidated convolutional multi-scale vision transformer based autism detection

Autism is a complex neurological condition characterized by repetitive behaviours, varied speech and difficulties in social interaction. Identifying autism spectrum using traditional methods alone is challenging due to its complexity and variability. In order to facilitate earlier and more accurate detection, Artificial intelligence (AI) methods have been developed to improve traditional diagnostic techniques. Standard diagnostic methods often struggle with reliability due to the variability and complexity of autistic individuals. To overcome these challenges, this study proposes a novel Deep Learning (DL) based framework for efficient and accurate autism spectrum detection. The input data are first pre-processed using an Adaptive Median Gaussian Filter (AMGF) to remove noise while preserving essential image features. Then, Generative Adversarial Networks (GANs) are employed for data augmentation, effectively overcoming dataset imbalance and security. Feature extraction is carried out using the Deep Multi-scale and multi-level enclosed attentional DarkNet (DMM-AND) model, which captures both fine-grained and high-level patterns. Finally, autism spectrum detection is performed using the Optimized Quadrangle Attention consolidated Convolutional Multi-scale Vision Transformer (OQA-CMVT) model. Here, the Random Spiral exploit chimp optimization algorithm (RS-COA) is used for the hyperparameter tuning process. The proposed technique achieves 97.96% accuracy for the autistic children's facial dataset and 98.04% for the Autism_Image_Data dataset.