A website-based Skin Disease Identification using a Convolutional Neural Network for Childcare Applications

Abstract Skin conditions in children, including chickenpox, hand-foot-mouth disease (HFMD), heat rash, and herpes zoster, remain diagnostic difficulties in many developing regions, including the Middle Eastern countries, especially in childcare environments. Childcare professionals frequently face challenges in accurately recognizing these conditions using solely observational methods. This initiative aims to fill this void by developing a web-based tool. The website utilizes transfer learning with the InceptionV3 convolutional neural network (CNN) model. It enables non-physicians to perform visual assessments and deliver prompt diagnoses with greater ease. The study was organized into three clear phases: Training, Website Development, and Testing. The 520-image dataset was partitioned into training (80%), validation (10%), and testing (10%) subsets to facilitate thorough model training and assessment. The model achieved an accuracy of 88% in classifying paediatric skin diseases. A nested 5-fold cross-validation (CV) strategy was adopted under three different experimental settings: with augmentation, without augmentation, and with a 50% dataset using augmentation, to further evaluate the model’s performance. The results showed that a nested 5-fold CV with augmentation achieved 0.7538 ± 0.0703 accuracy (mean ± standard deviation across folds), with a 95% bootstrap confidence interval of 0.700–0.804 (mean = 0.756). This promising accuracy suggests that the model can offer valuable support in diagnosing conditions that are typically challenging to detect through manual observation. The goal of this investigation is to minimize the transmission of infections in childcare settings by providing a precise detection instrument for skin ailments. Rather than replacing clinical expertise, the website aims to act as a decision-support tool for childcare workers, enabling timely response and reducing infection risks. Nonetheless, the relatively small dataset remains a limitation, emphasizing the need for future validation on larger, clinically verified datasets.