A Master thesis was discussed in Department of Computer Engineering/College of Engineering at University of Mosul entitled “Non-Invasive Infant Jaundice Level Estimation from Images of the Skin Using Machine Learning” submitted by the student (Banan Khalid Abdulkader Al-Dabbagh) on Wednesday, Feb. 19, 2025.

The study investigates a non-invasive, image-based technique for detecting and classifying neonatal jaundice severity using advanced Machine Learning (ML) technique. A new dataset of 344 images of infants with Jaundice and healthy of full-term newborns were created with four sub-datasets for classification. The research employs Deep Transfer Learning (DTL) with pre-trained models (VGG16, ResNet50, EfficientNet) and the K-Nearest Neighbors (KNN) algorithm. It also uses handcrafted techniques for feature extraction from scratch.
The study introduces two to five class classifications, with the five-class approach being novel. Several experiments and scenarios were conducted to optimize the process, demonstrating that DTL-based methods achieved higher accuracy, reaching up to 97.13%, outperforming the handcrafted method and some state-of-the-art. This research highlights the potential of DTL for improving early jaundice detection and reducing infant mortality, particularly in resource-limited settings. The methodology could lead to early jaundice detection solutions for home use alongside clinical applications.