Automatic deep learning method for analysis and prediction of neonatal hyperbilirubinemia in magnetic resonance imaging.

BACKGROUND

This study conducted a comparative analysis among newborns with varying levels of hyperbilirubinemia, explored the relationships between magnetic resonance imaging (MRI) image features and serum bilirubin levels in hyperbilirubinemia, and proposed an automatic classification system based on deep learning (DL) for prediction of neonatal hyperbilirubinemia (NHB).

METHODS

This retrospective study enrolled 606 consecutive neonates who had their serum bilirubin detected at the Xi'an Fourth Hospital, including 273 cases of patients and 333 cases of normal controls. After data preprocessing, MRI images were fed into the Inception-v3 network, graph convolutional network (GCN), and 3-dimensional (3D) patch-based GCN that introduced the graph attention mechanism (our GCN) for NHB analysis and classification, respectively. Multi-threshold grouping was conducted based on various serum bilirubin levels. Performance evaluation involved the area under the curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE).

RESULTS

As the bilirubin levels gradually increased, the overall performance metrics of DL system for detecting the T1-weighted imaging signal in the pallidum region showed a significant upward trend. Our GCN for the prediction and classification of MRI image features of NHB achieved satisfactory results. When the bilirubin value exceeded 400 µmol/L, it achieved an AUC of 0.86 and ACC of 0.81, which is significantly higher than other advanced models (ACC: 72-78.3%) with the same proposed input form.

CONCLUSIONS

The DL system has the potential to automatically analyze and predict NHB on MRI.