Non-Invasive Neonatal Jaundice Detection via Two-Phase Self-Supervised Learning and Vision Transformer

Neonatal jaundice is a potentially serious and prevalent condition that can lead to neurological challenges in newborns if not detected or treated on time. Currently, most approaches employed to diagnose neonatal jaundice are invasive and are often characterized by constraints in accuracy and accessibility, especially in locations with limited resources. To proffer a solution to these challenges, this study presents a vision transformer-based framework, JauNViT, for diagnosing neonatal jaundice. In this framework, a two-phase technique is deployed. The first phase involves the use of Vision Transformer (ViT) to extract deep features from the image data and then use the Self-Relative Clustering (SRC) technique to organize the image data into clinically meaningful clusters and learn discriminative features in the process, facilitating unsupervised learning from limited data. To further enhance the quality of these features, a self-supervised learning strategy based on Linear Centered Kernel Alignment (LCKA) loss was employed to refine the extracted representations. In the second phase, the pre-trained model was fine-tuned on a small labeled dataset, enabling the model to optimize its learned features for efficient neonatal jaundice classification. The proposed dual-phase approach based on ViT yields robust diagnostic accuracy and exhibits prospects of being adaptable to different healthcare settings.