Severe neonatal jaundice is a condition in newborns where high levels of bilirubin cause the skin and eyes to turn yellow, posing a risk of brain damage if not promptly treated. Current traditional methods are highly invasive and require intervention. Hence, this paper introduces a non-invasive approach for the preemptive detection of severe neonatal jaundice using computer vision and deep learning. The data processing pipeline includes image resizing, semantic segmentation, test split, data augmentation, and at last training-validation split. For this study, a custom CNN model was developed for binary classification alongside three transfer learning models to compare all four’s performance across key metrics such as accuracy, precision, recall, F1-score, and AUC. After training, all four models were saved and then used to classify a different dataset to evaluate their performance on visually distinct images. The Vision Transformer (1.23 GiB) and EfficientNet (320 MiB) models demonstrated superior performance on testing data, achieving AUC scores of 0.87 and 0.9, respectively. However, the custom CNN model (162 MiB) and Vision Transformer achieved 0.93 and 1.0 AUC score consistently on inference data, surpassing the other models. This research contributes to creating a contactless, frugal system that can be used as a mobile application, which predicts the chances of an infant having severe jaundice.

Biomedical Engineering Bilirubin Computer Vision Deep Learning Semantic Segmentation.