JCSE, vol. 18, no. 1, pp.1-9, 2024

DOI: http://dx.doi.org/10.5626/JCSE.2024.18.1.1

Classification of Autism Spectrum Disorder Based on Facial Images Using the VGG19 Algorithm

Wandi Yusuf Kurniawan and Putu Harry Gunawan
Human Centric Engineering (HUMIC), School of Computing, Telkom University, Bandung, Indonesia

Abstract: Autism spectrum disorder profoundly affects early communication and physical skills, emphasizing the need for effective interventions. Approximately one in a hundred children worldwide is affected by autism. The convolutional neural network (CNN), especially VGG19, is the most accurate tool for detecting autism using a facial image dataset. Notably, there are many configurations that can be applied to produce the best accuracy. This study evaluated how facial images can be used to classify autism using VGG19-based deep learning models with different configurations, such as longshort term memory (LSTM) and Dropout layers; adaptive moment estimation (Adam), root mean square propagation (RMSprop), and stochastic gradient descent (SGD) optimizers; and a cosine annealing learning rate scheduler. Results highlighted substantial performance variations across the configurations, with RMSprop+LSTM+Dropout achieving the highest accuracy (75.85%), average precision, non-autistic precision, and average F1-score. Notably, Adam showed the best performance in non-autistic precision (83.09%) and autistic F1-score (76.74%), while Adam+LSTM+Dropout demonstrated superior autistic precision (85.16%) and non-autistic recall (90.82%). Moreover, SGD+Dropout achieved the highest autistic recall (91.84%). Selecting an appropriate configuration is crucial, and further research can help optimize the architecture, activation functions, and preprocessing for enhanced accuracy. High-accuracy models hold promise for aiding autism detection and communication and physical skill development.

Keyword: Autism spectrum disorder; Facial image; CNN-VGG19; Classification; Optimizer

Full Paper:   11 Downloads, 45 View