Computer-Aided Diagnosis of Pneumoperitoneum on Neonatal Abdominal Radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neonatal Surgery

Background:

• Neonatal gastrointestinal perforation is a critical condition requiring prompt diagnosis.
• Interpreting neonatal abdominal radiographs for perforation is challenging.
• This study addresses the need for improved diagnostic tools in neonates.

Purpose of the Study:

• Develop a deep convolutional neural network (DCNN) model for pneumoperitoneum segmentation.
• Evaluate the DCNN's efficacy in assisting the detection of neonatal gastrointestinal perforation.

Main Methods:

• A multicenter retrospective study of 1,187 neonatal abdominal radiographs.
• A DeepLabV3+ model with ResNet50 backbone was fine-tuned for pixel-level segmentation.
• Pneumoperitoneum regions were annotated, and a pixel-based threshold was used for classification.

Main Results:

• The DCNN achieved a median Dice similarity coefficient of 0.81 for pneumoperitoneum segmentation.
• Segmentation performance correlated positively with pneumoperitoneum volume (Spearman ρ = 0.83).
• The model demonstrated excellent diagnostic accuracy (AUC, 0.999; sensitivity, 100%; specificity, 98.5%) for perforation detection.

Conclusions:

• The DCNN model shows robust performance in segmenting pneumoperitoneum and classifying gastrointestinal perforation.
• This AI tool has potential as a clinical decision support system for early neonatal gastrointestinal perforation detection.
• Further research is needed to validate generalizability and clinical integration.