Classifying Neonatal Acute Bilirubin Encephalopathy based on Spatial-Pathological Attention Module and Convolutional Neural Network

Neonatal hyperbilirubinemia is one of the most common causes of medical attention in the neonatal period. In severe cases, it can deteriorate into acute bilirubin encephalopathy, causing irreversible neurological damage. Due to the nonspecific nature of acute bilirubin encephalopathy, neonates with normal myelination and neonates with acute bilirubin encephalopathy exhibit similar hyperintensity on magnetic resonance imaging, making clinical diagnosis based on routine radiological readings difficult. As a deep learning framework, convolutional neural network has become the mainstream method of medical image classification in recent years. In this paper, an acute bilirubin encephalopathy classification strategy based on a convolutional neural network is proposed, which combines different acute bilirubin encephalopathy patterns with neonatal clinical phenotypic data. The baseline consists of three separate pathways used to train different acute bilirubin encephalopathy modes, namely T1, T2, and T2-Flair, and the feature maps of the different pathways are concatenated through the spatial-pathological attention module. A database of 800 patients is first screened by combining their clinical hemoglobin concentrations to produce a new dataset of 650 patients. Each sample consists of three morphologies of 3D brain acute bilirubin encephalopathy and their corresponding clinical phenotypic data. The superiority of the proposed method is proved by designing a cautery experiment and exploring different classification strategies.