Detecting autism with Vietnamese child facial images using deep learning

Deep learning techniques created a significant increase in intelligent systems, especially in the medical field. Among mental problems, autism is a dangerous neurodevelopmental disorder and it needs to be diagnosed early because of the malleability of child brain development. In our study, we focused on autism detection by using the Vietnamese facial child image and studied the role of international data and Vietnamese data when applying deep learning approach to diagnose autism. To do that, we proposed different strategies based on our hypothesis about factors of the transfer learning and training set types. To conduct the experiment, we prepared a Vietnamese facial child image set from several kindergartens in Ho Chi Minh City, Vietnam and we applied different deep architectures such as ResNet, DenseNet, and AlexNet in the autism classification experiment with both Vietnamese and international facial child images. We analyzed important factors from the experiment results with area under the curve (AUC), accuracy, sensitivity, and specificity, including applying transfer learning and the appearance of Vietnamese data in the training set. Besides, we also discussed the difference of international and Vietnamese data domains. The exposure of data distribution differences in the proposed strategies also highlights the importance of collecting facial data of Vietnamese children.