Computer-Aided Diagnosis of Pneumoperitoneum on Neonatal Abdominal Radiographs

Neonatal gastrointestinal perforation is a life-threatening condition that requires timely and accurate diagnosis. However, interpreting abdominal radiographs in this population is often challenging. In this study, we aimed to develop a deep convolutional neural network (DCNN) model to segment pneumoperitoneum on neonatal abdominal radiographs and to evaluate its potential to assist in detecting neonatal gastrointestinal perforation.

This multicenter retrospective study included 1,187 abdominal radiographs (181 perforation and 1,006 control images) from neonates with gastrointestinal perforation and controls. Pneumoperitoneum regions were annotated by experienced clinicians. The dataset was randomly divided into training (n = 830), validation (n = 118), and test (n = 239) sets. A DeepLabV3+ model with ResNet50 backbone was fine-tuned for pixel-level segmentation. A single pixel-based threshold, derived from ROC analysis, was used to classify gastrointestinal perforation, with diagnostic performance subsequently compared to that of clinicians.

The DCNN model achieved a median Dice similarity coefficient of 0.81 on the test dataset, indicating strong overlap between predicted and actual pneumoperitoneum regions. Furthermore, segmentation performance was positively correlated with pneumoperitoneum volume (Spearman ρ = 0.83, p < 0.001). Classification using the pixel-based cut-off demonstrated excellent diagnostic accuracy (AUC, 0.999; sensitivity, 100%; specificity, 98.5%), comparable to experienced clinicians.

The DCNN model demonstrated robust segmentation and classification performance, highlighting its potential as a clinical decision support tool for early detection of gastrointestinal perforation in neonates. Future studies should validate the model's generalizability and assess its integration into clinical practice.