A Fusion Framework for Early Autism Identification Using Transformer-Based Deep Learning Approach

Autism Spectrum Disorder (ASD) is a disorder that impacts brain development and can influence an individual’s perceptions and social interactions. Communication and social interactions can become challenging as a result. More than ever, autism spectrum disorder is becoming more common. Autism can be identified at any developmental stage, and symptoms typically start to show up between the ages of two and four. Compared to children who are typically developing, those with ASD exhibit distinct facial landmarks. In order to apply appropriate treatments and enhance the quality of life of individuals with autism, early detection is crucial.Numerous indicators, including genetic screening, brain MRI, and facial images, can be used to identify autism. However, one of the main difficulties in diagnosing autism is that its symptoms can be confused with those of many other conditions. In this study, we proposed a deep learning-based fusion framework with MobileNetV1, which is suitable for mobile phones, and the use of the transformer model to improve their classification performance. The experimental results showed that the transformer model significantly improves the classification performance, and the fusion model has better classification performance than any of the classifiers participating in the study. The model was evaluated in terms of accuracy, confusion metrics, precision, sensitivity, specificity, F1 score, error rate, AUC, and G mean. The proposed model has an accuracy of 91.33%, with an error rate of 0.0866, which is lower than the error rate of the previous studies.