Autism Spectrum Disorder Detection using Deep Learning Techniques

In recent years autism spectrum disorder (ASD) has been rapidly increasing in people irrespective of age which effects social interactions, communication and behaviour in daily life. Early diagnosis is essential for effective intervention; however, existing diagnostic procedures rely heavily on behavioural assessments, which are subjective, time-consuming, and inaccessible in under-resourced settings. In this research deep learning–based diagnostic framework using facial imagery is proposed to classify ASD and non-ASD cases. The system combines two pretrained convolutional neural networks, VGG16 and InceptionV3, into a weighted soft-voting ensemble. To address dataset imbalance, SMOTE is applied during training, and threshold tuning is used to optimize classification boundaries based on F1-score. The proposed model achieved 97.11% accuracy and AUC of 0.99 on a curated facial image dataset. Additionally, Grad-CAM is used to enhance interpretability by highlighting salient facial regions. The pipeline is fully reproducible via Google Collab with GPU support. The results confirm the ensemble’s effectiveness as a scalable, non-invasive screening tool for early ASD detection.