Automatic Non-Invasive Jaundice Detection of Neonates Using Computer Vision

Smartphone-based screening for neonatal jaundice offers a low-cost alternative to conventional bilirubin measurement. Still, most computer vision techniques rely on color calibration, fixed skin segmentation, or a limited set of features, which reduces their reliability across different imaging conditions. In this work, we propose a calibration-free, smartphone-based approach for assessing neonatal jaundice by combining adaptive skin segmentation, multi-color feature extraction, SHAP-guided feature selection, and machine learning. Our framework uses visual features from the images to direct the process to either SAM or DensePose to extract the skin from public images or more complex clinical images. The skin areas are enhanced and represented in the RGB, HSV, CIE Lab, and YCbCr color spaces, and the chromatic features are then ranked and selected using SHAP. The selected features were evaluated using machine learning algorithms to estimate TsB continuously and to perform three-class jaundice classification. The experiments were carried out on the available NeoJaundice dataset and a clinical NICU dataset collected at Benha University Hospital. The segmentation evaluation supported adaptive segmentation, with Dice/IoU scores of 0.8807/0.7903 for DensePose (Benha) and 0.9850/0.9705 for SAM (NeoJaundice). For TSB estimation, KNN obtained an R of 0.982, an R2 of 0.965, and an RMSE of 0.973 mg/dL. For the three-class classification, Random Forest achieved accuracies of 98.31% on the NeoJaundice dataset and 98.44% on the clinical dataset. When evaluated on 261 clinical images not previously seen, an accuracy of 98.47% was achieved, providing evidence of its performance on previously unseen NICU images. The results show the promise of integrating adaptive segmentation, compact multi-color representations, and explainable feature selection for calibration-free smartphone-based screening of neonatal jaundice. Further multicenter validation across diverse neonatal populations, skin tones, and smartphone devices is needed to establish broader clinical applicability.