Deep Learning-Based Autism Detection Using Facial Images with Hybrid Preprocessing Techniques

Autism Spectrum Disorder (ASD), a neurodevelopmental disorder, presents with impairments in socialization and communication skills. The intervention of the condition largely depends on the early detection of the disorder. However, the traditional detection methods of the disorder are largely subjective in nature and require a considerable amount of time to be spent. In the following paper, a deep learning approach to detect an automated method of detecting autism by analyzing facial images is proposed. The image dataset contains 2,936 facial images, equally divided into two classes of autistic and non-autistic images. The preprocessing steps involved are explained in detail, and some of them are basic filters such as bilateral filtering, CLAHE (Contrast Limited Adaptive Histogram Equalization), gamma correction, and denoising. Hybrid combinations of filters are also used, such as the combination of bilateral filtering and CLAHE, as well as unsharp filtering. These steps are meant to enrich feature presentation and consequently enhance classification effectiveness. All images are downsized to 128x 128x 3, normalized, and data are augmented to enhance model generalizability. Five deep learning models, including CNN, InceptionV3, MobileNetV2, VGG16, and DenseNet are tested in the conditions of raw and pre-processed images. The results of the experiment indicate that preprocessing has a tremendous impact on the performance of the model. Without preprocessing, MobileNetV2 has the highest accuracy of 79.91 %; with preprocessing, DenseNet achieves the best testing accuracy of 79.00 % and F1-score of 0.79. The results show that hybrid preprocessing methods significantly improve classification performance, whose extent of improvement is model dependent. The suggested framework is therefore an effective and non-invasive approach to detecting early autism through the use of facial images.