Computer Vision-Based Neonatal Jaundice Detection Using Transfer Learning: A Three-Architecture Comparative Study

Abstract

Neonatal jaundice (neonatal hyperbilirubinemia) is one of the most common conditions of early infancy, resulting from increased bilirubin levels giving rise to yellow skin and yellow sclera. While most cases are physiological and will pass, failure to recognize severe hyperbilirubinemia may result in bilirubin-induced neurological dysfunction, acute bilirubin encephalopathy, and kernicterus. The most common laboratory tests for the diagnosis are total serum bilirubin and, in many settings, transcutaneous bilirubinometry. However, these can be invasive, expensive, device-dependent, or not readily available repeatedly in low-resource settings. Photographic image analysis is a potential method for low-cost, non-invasive screening as jaundice becomes visible. In this work, we introduce a deep learning model for binary neonatal jaundice detection using photographs, called NeoJaundice AI. Further, we apply three different convolutional neural network architectures (EfficientNetB3, Modified ResNet50V3, and MobileNetV2). We employed a transfer learning procedure via stratified dataset splitting, two-phase training, label smoothing, and receiver operating characteristic (ROC)-based threshold calibration to train the models. The base model was EfficientNetB3, achieving 87.2% test set accuracy and a ROC-AUC of 0.9570. The next model we tested, the modified ResNet50V3, achieved test accuracy (94.29%) and ROC-AUC (0.9792) with the highest recall for jaundice. The third model, MobileNetV2, performed best overall with a test accuracy of 96.0%, ROC-AUC of 0.9883, and balanced precision, recall, and F1-score of 0.96 for the jaundice and normal classes. Based on these results, it is concluded that MobileNetV2, has the best performance in terms of diagnostic accuracy, computational efficiency, and deployment. The proposed NeoJaundice AI system thus paves the way for the feasibility of AI-aided image-based neonatal jaundice screening, especially for point-of-care applications at the edge and in mobile environments that require quick, non-invasive triage.

Keywords: Neonatal jaundice; neonatal hyperbilirubinemia; bilirubin detection; deep learning; convolutional neural networks; transfer learning; mobile health; edge deployment.