Towards Robust Non-Invasive Neonatal Jaundice Detection: Problem Formulation and Analysis of Real-World Challenges

Abstract

Neonatal jaundice (neonatal hyperbilirubinemia) appears in 60–80% of newborns around the world and is one of the leading causes of neonatal death and of neurological disorders due to a lack of treatment in many end, and even some developing, countries, especially of those in sub-Saharan Africa and South Asia. For the majority of laboratories, the gold standard for diagnosis is invasive testing of serum bilirubin. It is painful and expensive and is often not an option in developing countries. Clinical research on non-invasive testing, such as the use of transcutaneous bilirubinometry (TcB) or the use of deep learning systems and imaging-based systems, is promising. However, these techniques tend to be limited by five interrelated practical challenges: (1) the effect of melanin and the associated skin tone bias, and the resulting inequity in diagnosis across the different levels of the Fitzpatrick scale, (2) the small and homogeneous datasets, as the majority of the published deep learning solutions have been trained on fewer than 300 neonates from a specific geographic location, (3) the systems are not robust to variations in real-life imaging conditions such as different lighting, cameras or imaging strategies, (4) all systems describe a lack of research that validates the internal imaging and segmentation of the body, and (5) a lack of thorough research in different imaging and segmentation methods. This article is a systematic analysis of all records currently published. We identify the limitations of current systems, and in relation to those failures, we assign performance tiers. We have created PPC (Perceptual Pre-processing Correction) as a framework to solve the issues, and from that framework, we created new focused research questions. The framework contains several original inventions, including the design of a biased skin tone melanin color norm spotting Convolutional Neural Network (ST-CNN) method, the CZZBP (Colour-Zone-based Z-score Bilirubin Prediction) segmentation method, and the melanin-informed color normalization method.