Neonatal jaundice detection using a vision transformer-based deep learning model

Abstract

Neonatal jaundice is a prevalent and potentially serious condition that can lead to severe complications if undiagnosed or untreated. While traditional diagnostic methods like blood sampling are invasive and time-consuming, and transcutaneous bilirubinometers remain costly, smartphone-based image analysis offers a promising low-cost, non-invasive alternative. However, most existing solutions rely on traditional machine learning techniques with limited accuracy and generalizability. In this study, we introduce a deep learning approach based on the Vision Transformer (T2T-ViT) and compare its performance with three other models, ResNet, Support Vector Machine (SVM), and K-Nearest Neighbors (k-NN), using a clinically annotated dataset of neonatal skin images captured via a smartphone camera. The models were evaluated using multiple performance metrics including accuracy, precision, recall, F1-score, Matthews Correlation Coefficient (MCC), and Area under the Curve (AUC). The T2T-ViT model achieved 99% across all metrics, significantly outperforming both convolutional and traditional machine learning models. These findings demonstrate the feasibility of applying transformer-based deep learning architectures for accessible, scalable, and accurate non-invasive neonatal jaundice screening, potentially enabling early intervention in resource-limited settings. This approach could serve as an accessible, scalable screening tool for neonatal jaundice detection, particularly in low-resource clinical settings.