YOLO–SAM-Guided ROI-Based Deep Learning for Non-Invasive Neonatal Jaundice Detection

Early and reliable detection of neonatal jaundice is of critical importance for preventing bilirubin-induced neurological damage. However, visual assessment methods widely used in clinical practice are subjective and highly dependent on the examiner and environmental conditions. In this study, three different deep learning–based approaches were systematically evaluated for the automatic detection of neonatal jaundice using the NeoJaundice dataset. In the first approach, direct Convolutional Neural Network (CNN)-based classification was performed on raw images using the ResNet18 and MobileNetV2 architectures. In the second approach, automatic pixel-level region of interest (ROI) extraction was applied prior to classification using YOLO-based coarse localization and the Segment Anything Model (SAM). In the third and most advanced approach, CNN features ex-tracted from the ROI were transferred to an Long Short-Term Memory (LSTM) network with the aim of modeling contextual dependencies between spatial representations. To the best of our knowledge, this study is the first to systematically examine the effect of YOLO–SAM–based automatic ROI extraction on classification performance using a color-sensitive and clinically labeled neonatal jaundice dataset. Experimental results show that ROI-based models improve inter-class balance and sensitivity, while the hybrid CNN–LSTM architecture achieves the highest overall accuracy (76.45%) when combined with ResNet18 features. The obtained findings demonstrate that combining precise ROI extraction with contextual feature modeling provides a more reliable and generalizable framework for camera-based neonatal jaundice detection systems, particularly in low-resource and non-invasive clinical screening scenarios.