Improved Deep Learning Model for Prediction of Dermatitis in Infants

Abstract

Indonesia's equatorial climate, characterized by summer and rainy seasons, presents environmental conditions that contribute to a high incidence of dermatitis in infants. Dermatitis, an inflammatory skin condition, can lead to significant discomfort in infants, affecting their sleep, growth, and development. Early diagnosis is crucial for effective treatment; however, conventional diagnostic methods in clinics and hospitals—such as physical observation and parental interviews—are often time-consuming, subjective, and may lack precision, creating a need for more efficient diagnostic tools. This study explores the application of deep learning models to enhance the accuracy and speed of dermatitis diagnosis in infants. Four convolutional neural network (CNN) models were evaluated: MobileNet, VGG16, ResNet, and a Custom CNN model specifically designed for this study. Using a dataset of 1,088 skin images collected from three regions in Riau Province, Indonesia, we conducted training and testing to assess each model’s performance in distinguishing between dermatitis-affected and healthy skin. Results show that MobileNet and the Custom CNN outperformed other models, achieving accuracy rates of 97% and 85%, respectively. MobileNet’s high accuracy and efficiency make it a viable option for mobile applications, enabling rapid, on-site diagnosis in resource-limited settings. The Custom CNN model, tailored to the unique features of infant skin, also showed promising results. These findings demonstrate the potential of automated, image-based diagnostic tools for assisting medical professionals in early dermatitis detection, improving patient outcomes. This study contributes a valuable diagnostic solution that leverages deep learning to support healthcare providers, particularly in areas with limited access to specialized medical resources.