This study presents a deep learning approach for the early diagnosis of autism spectrum disorder (ASD) in children by analyzing facial features. The research aims to detect autism using a deep learning model trained on a dataset of facial images from children with and without ASD. Three pre-trained convolutional neural network (CNN) models—VGG16, VGG19, and EfficientNetB0—were used as feature extractors and binary classifiers. The dataset, sourced from Kaggle, contains 3014 images of children, with 2536 images used for training, 300 for testing, and 100 for validation. The models achieved accuracies of 84.66%, 80.05%, and 87.9%, respectively. The study highlights the potential of deep learning, particularly CNNs, in diagnosing ASD through facial features. CNNs are effective for image classification due to their ability to learn from large datasets. Transfer learning was employed to adapt pre-trained models to the specific task of ASD detection, which is more efficient than training models from scratch. The research also discusses the importance of data augmentation and preprocessing to enhance model performance. The study contributes to the field by demonstrating the effectiveness of using pre-trained models and transfer learning for ASD diagnosis. It also emphasizes the need for models with minimal hyperparameters to ensure efficiency and adaptability to varying dataset sizes. The results indicate that EfficientNetB0, when trained with the Adamax optimizer and a learning rate of 0.001, achieved the highest accuracy of 88.33% and an AUC of 95.44%. The study concludes that deep learning techniques, particularly CNNs, offer a promising approach for early ASD diagnosis through facial analysis.This study presents a deep learning approach for the early diagnosis of autism spectrum disorder (ASD) in children by analyzing facial features. The research aims to detect autism using a deep learning model trained on a dataset of facial images from children with and without ASD. Three pre-trained convolutional neural network (CNN) models—VGG16, VGG19, and EfficientNetB0—were used as feature extractors and binary classifiers. The dataset, sourced from Kaggle, contains 3014 images of children, with 2536 images used for training, 300 for testing, and 100 for validation. The models achieved accuracies of 84.66%, 80.05%, and 87.9%, respectively. The study highlights the potential of deep learning, particularly CNNs, in diagnosing ASD through facial features. CNNs are effective for image classification due to their ability to learn from large datasets. Transfer learning was employed to adapt pre-trained models to the specific task of ASD detection, which is more efficient than training models from scratch. The research also discusses the importance of data augmentation and preprocessing to enhance model performance. The study contributes to the field by demonstrating the effectiveness of using pre-trained models and transfer learning for ASD diagnosis. It also emphasizes the need for models with minimal hyperparameters to ensure efficiency and adaptability to varying dataset sizes. The results indicate that EfficientNetB0, when trained with the Adamax optimizer and a learning rate of 0.001, achieved the highest accuracy of 88.33% and an AUC of 95.44%. The study concludes that deep learning techniques, particularly CNNs, offer a promising approach for early ASD diagnosis through facial analysis.