Virtual Booth

ABSTRACT

Background: Kawasaki disease (KD) is deadly as it causes the blood vessels to become inflamed and swollen, which can lead to complications in the blood vessels that supply blood to the heart. Without treatment, approximately 1 in 4 children with Kawasaki disease get heart complications. The symptoms, however, are like hand-foot-and-mouth disease (HFMD) which is milder, but contagious viral infection common in young children. Both diseases share common early clinical signs such as fever, and rash. Problem: Due to the similarity in the early-stage symptoms, parents may be misled by assuming HFMD when the fact that the child has KD. The delay in treatment would then lead to serious repercussions that can result in life-threatening complications. Solution: To solve this problem, the project has developed an intelligent system based on deep learning that can detect and differentiate KD from HFMD based on skin images and AlexNet convolutional neural network (CNN). Methodology: A total of 707 image samples for KD, HFMD, and healthy control has been acquired from internet database. The samples underwent preprocessing stage to standardize the image dimensions with minimal information loss. Augmentation is then performed to enhance the size of the dataset and improve robustness of the system against translational invariance. The samples are randomized and split for training and validation with split ratio of 80:20. The dataset is then used to train the AlexNet architecture with adaptive moment estimation (Adam) optimizer, batch size of 512 and learning rate of 0.0001. Outcomes: The network is successfully developed to classify skin images of KD, HFMD, and healthy control with accuracies of 99.8% for training, and 99.6% for validation. Potential: While the system is still at prototype stage, the intelligent model can be enhanced to identify other diseases with symptoms emanating through the skin and miniaturized into mobile application.