Leveraging convolutional neural networks for precise diagnosis of autism through transfer learning and ensemble model

Autism is a neurodevelopmental condition that can make social interaction and communication challenging. It is known for repetitive behaviors and difficulties in understanding nonverbal communication cues. The growing interest in autism, fueled by advances in global health expertise, has resulted in increased efforts to enable early detection and intervention. The fundamental purpose of early detection is to provide personalized behavioral development treatment programs, easing the integration of autistic people into the society. In this research, we investigate the efficacy of deep learning models for autism detection, by developing an ensemble of high-performing Convolutional Neural Network(CNN)-based models. We perform a comparative study of five deep learning models, Xception, InceptionV3, MobileNetV2, DenseNet201, ResNet50V2 to analyze images of autistic and non-autistic children. To enhance the predictive performance and robustness, we create an ensemble model combining the three top-performing deep learning models, achieving a state-of-the-art 91.7% accuracy. We further improve our performance to 93.33% by performing histogram equalization on the images. The use of automatic hyperparameter tuning has helped explore alternative model configurations effectively and resulted in a highly optimized CNN architecture for autism diagnosis in children. Our ensemble model achieved state-of-the-art results using a publicly available dataset, which includes 2 classes and comprises a total of 2940 images for training, validation, and testing. Our research showcases the immense potential of deep learning and ensemble model in advancing autism detection and early intervention.