Spatio-Temporal Graph Convolutional Network on Sequential Chest X-Rays for Early Bronchopulmonary Dysplasia Prediction

One of the most prevalent and harmful respiratory conditions in preterm infants is bronchopulmonary dysplasia (BPD), particularly those born very early and with extremely low birth weight. In the past, BPD was believed to be a chronic progressive lung disease that also affected the pulmonary parenchyma and vascular abnormalities. Children with BPD are much more likely to experience long-term cognitive dysfunction, cerebral palsy, speech impairment, pulmonary dysfunction, hearing and visual impairment, such as retinopathy of prematurity, and other conditions than children without BPD. Physicians used chest radiographs and their subjective judgment to diagnose BPD. Different subjective evaluations of BPD may result from physicians' varying clinical experiences and physical circumstances. In order to improve the quality of medical care and enable physicians to make more accurate and early diagnoses, a machine learning (ML)-based system is employed for the effective and efficient prediction of lung development in preterm newborns. This study aims to develop a Spatio-Temporal Graph Convolutional Network on Sequential Chest X-Rays for Early Bronchopulmonary Dysplasia Prediction (STGCN-CEBDP). At first, data preprocessing is used to pre-process the raw information from chest X-ray imaging. Next, DenseNet-3D is applied for the feature extraction process to capture both spatial patterns in chest X-ray sequences and temporal changes over time, which can enhance early BPD prediction. Following, STGCN is employed for the BPD prediction process. Finally, Whale Optimization Algorithm (WOA)-based hyperparameter tuning is conducted to enhance outcomes of accurate STGCN prediction. The simulation outcomes of the STGCN-CEBDP approach is tested on medical imaging database, and the experimental analysis demonstrate the enhancements of the STGCN-CEBDP method over other DL models.