The National Center for Child Health and Development (NCCHD) has taken the opportunity of the adoption of the AI Hospital Project to conduct data science awareness and education activities not only for researchers and doctors but also for all employees. As one of the training activities, we created a face photo data set to build an AI using supervised machine learning.

Although it is relatively easy to collect facial images via the Internet, it is a challenging task because it is desirable to have a fixed posture for medical and biological applications, and personal information must be protected.

However, it became big data with labels for gender and smiles on an unprecedented scale, and by using data expansion and convolutional neural networks utilizing transition learning, we were able to present a model showing recognition accuracy of 98.2% for gender and 93.0% for smiles. Many congenital diseases in children show characteristic facial features, and there are high expectations for diagnosis by image recognition. However, the number of patients with these rare diseases is small, and the large amount of data required for implementation by deep learning has been an obstacle.

The results of this study will be used to examine the amount of data required for AI construction and provide guidance for future AI development.