Artificial intelligence aids doctors in diagnosing necrotizing enterocolitis and predicting surgery using abdominal radiographs: a multicenter study.

Background: Neonatal necrotizing enterocolitis (NEC) is challenging to diagnose due to its subtle radiological features on abdominal radiographs (ARs) and considerable variability in interpretation among clinicians, especially those with limited experience, which may delay timely intervention. The study aimed to develop a convolutional neural network (CNN) based artificial intelligence (AI) model using ARs to predict NEC and the need for surgical intervention, and to evaluate its ability to assist clinicians in AR interpretation. Methods: We retrospectively collected 738 ARs from 576 preterm infants across three medical centers. The ARs were labeled into three categories: non-NEC, medical-NEC, and surgical-NEC. Data from two centers (set A, n=347; set B, n=130) were randomly split into training and internal testing sets (8:2 ratio), while data from the third center (set C, n=99) served as an external testing set. Six deep learning models [efficientnet, Inception_v3, visual geometry group (VGG), resnet, squeezenet, and densenet] were trained, and the best model was selected. Six clinicians with varying experience interpreted set C ARs, and their accuracy was compared before and after using the best model's AI-assisted visualizations. Results: The Efficientnet-b0 model, trained on the combined dataset of A and B, performed best. On set C, it achieved area under the receiver operating characteristic curve (AUC) values of 0.883, 0.640, and 0.837 for non-NEC, medical-NEC, and surgical-NEC, respectively, outperforming single-center models (P<0.05). With AI assistance, clinicians' diagnostic accuracy improved by 2.0% to 26.2%. Conclusions: The CNN-based AI model effectively distinguished NEC and identified cases requiring surgery. Clinicians' AR interpretation accuracy improved with AI assistance.