Neonatal intestinal perforation, a life-threatening condition where holes form in the intestines, has been difficult to diagnose early because X-ray scans often fail to clearly show the characteristic “air in the abdominal cavity” sign.

However, artificial intelligence (AI) can now fill this gap.

Intestinal perforation, which primarily affects premature infants, can lead to complications or even death if diagnosis is delayed. Yet in neonatal intensive care units (NICUs), it is often difficult for radiologists to interpret images immediately, posing a high risk of delayed or incorrect diagnosis.

Asan Medical Center (AMC) announced on Wednesday that a research team—led by Professor Yoon Hee-mang of the Department of Radiology, Professor Kim Nam-kug of the Department of Convergence Medicine, and Professor Lee Byong-sop of the Department of Neonatology—has developed an AI interpretation model.

This model analyzes neonatal X-ray images using artificial intelligence to determine the presence of intestinal perforation and even locate the lesion.

The AI model demonstrated remarkable performance, achieving an internal validation accuracy of 94.9 percent and an external validation accuracy of 84.1 percent.

The team built a deep multi-task learning model that classifies the presence of intestinal perforation using neonatal X-ray images while simultaneously learning to identify and mark areas filled with air within the abdominal cavity. Of approximately 2.6 million pediatric X-ray images collected at AMC from January 1995 to August 2018, 294 images showing intestinal perforation and 252 control images were selected for training.

To better capture the varied imaging patterns of bowel perforation across patients, the researchers applied data augmentation techniques, enabling the model to recognize diverse manifestations and lesion sites. For external validation, they secured 64,000 images from 11 domestic hospitals, ultimately selecting 164 intestinal perforation images and 214 control images for multi-center testing.

Results showed the AI model achieved a diagnostic accuracy of 94.9 percent in internal validation, accurately identifying intra-abdominal air, while external validation confirmed 84.1 percent accuracy—comparable to specialist performance.

When medical staff used the AI tool as an aid, diagnostic accuracy improved from 82.5 percent to 86.6 percent, while inter-reader agreement significantly increased from 71 percent to 86 percent.

“Neonatal intestinal perforation is highly urgent, making rapid diagnosis paramount. But imaging findings are ambiguous and differ from adults, meaning diagnostic accuracy depends heavily on reader experience,” Professor Yoon said. “This AI model not only matched specialist-level accuracy but also improved consistency among medical staff.”

Professor Kim added, “We are focusing on technologies critically needed in clinical practice but still under-researched, such as neonatal intestinal perforation. By developing and applying AI models that support early diagnosis in NICUs, we aim to contribute to improved neonatal survival rates.”