Explainable Deep Learning for Neonatal Jaundice Classification Using Uncalibrated Smartphone Images

Hyperbilirubinemia, commonly known as jaundice, is a prevalent condition in newborns, primarily arising from alterations in red blood cell metabolism during the first week of life. While conventional diagnostic methods, such as serum analysis and transcutaneous bilirubinometry, are effective, there remains a critical need for robust, non-invasive, image-based diagnostic tools. In this study, we propose a custom-designed convolutional neural network for classifying jaundice in neonatal images. Image preprocessing and segmentation techniques were systematically evaluated. The optimal workflow, which incorporated contrast enhancement and the extraction of regular skin patches of 144 × 144 pixels from regions of interest segmented using the Segment Anything Model, achieved a testing F1-score of 0.80. Beyond performance, this study addresses numerous shortcomings in the existing literature in this area relating to trust, replicability, and transparency. To this end, we employ fair performance metrics that are more robust to class imbalance, a transparent workflow, share source code, and use Gradient-weighted Class Activation Mapping to visualise and quantify the image regions that influence the classifier’s predictions in pursuit of epistemic justification.