Asan Medical Center Develops AI Interpretation Model to Detect Presence of Air in Newborns’ Abdominal Cavities and Identify Lesion Sites

Model Analyzes X-ray Images Alone, Showing High Clinical Applicability… Internal Validation Accuracy Reaches 94.9%

“Supports Early Diagnosis in Neonatal Intensive Care Units Where Rapid Decisions Are Critical, Contributing to Improved Survival Rates of Newborns”

▲ (From left) Professor Hee Mang Yoon of the Department of Radiology, Professor Namkug Kim of the Department of Convergence Medicine, and Professor Byong Sop Lee of the Division of Neonatology at Asan Medical Center 

Intestinal perforation in newborns is a life-threatening condition in which a hole forms in the intestine due to diseases such as necrotizing enterocolitis. Diagnosis is typically made by checking X-ray images for the presence of air in the abdominal cavity, but the signs of perforation are often subtle, making accurate interpretation challenging.

A research team led by Professors Hee Mang Yoon of the Department of Radiology, Namkug Kim of the Department of Convergence Medicine, and Byong Sop Lee of the Division of Neonatology at Asan Medical Center recently announced that they have developed an artificial intelligence (AI) interpretation model capable of analyzing newborns’ X-ray images to determine the presence of intestinal perforation and pinpoint the location of the lesion.

The AI interpretation model for neonatal intestinal perforation demonstrated high performance, with an internal validation accuracy of 94.9% and an external validation accuracy of 84.1%. In particular, its ability to detect intestinal perforation at an early stage is expected to help improving the survival rate of newborns.

Intestinal perforation, which occurs primarily in premature infants, can lead to severe complications or even death if diagnosis is delayed. However, due to the nature of neonatal intensive care units, it is often difficult for radiologists to interpret images immediately, increasing the risk of misdiagnosis or delayed diagnosis.

In addition, existing AI interpretation models have been developed based on adult data, making them difficult to apply to newborns, who differ greatly from adults in body size, X-ray positioning, and radiologic findings.

To address this issue, the research team developed a deep multitask learning model that classifies the presence of intestinal perforation using newborns’ X-ray images while simultaneously identifying and highlighting areas in the abdominal cavity filled with air.

To enable the AI model to recognize various patterns and lesion locations of intestinal perforation, which can differ from patient to patient, the research team applied data augmentation techniques, allowing the model to effectively learn from a wide range of imaging variations.

To verify the model’s clinical applicability, the research team collected 64,000 X-ray images from 11 hospitals nationwide as external validation data. Among these, 164 images showing intestinal perforation and 214 control images were selected for internal and multi-center external validation.

In the internal validation, the AI interpretation model for intestinal perforation achieved a diagnostic accuracy of 94.9%, accurately identifying the areas in the abdominal cavity filled with air.

When validated with external data, the model achieved a diagnostic accuracy of 84.1%, comparable to that of medical specialists. The research team also assessed the assistive effect of using the AI interpretation model and found that diagnostic accuracy improved from 82.5% to 86.6%. Notably, inter-observer agreement increased significantly from 71% to 86%.

Professor Hee Mang Yoon of the Department of Radiology at Asan Medical Center explained, “Neonatal intestinal perforation is a medical emergency where rapid diagnosis is crucial, but the imaging findings are often subtle and differ from those of adults, making diagnostic accuracy highly dependent on the radiologist’s experience. The AI interpretation model for neonatal intestinal perforation not only demonstrated specialist-level accuracy but also improved diagnostic consistency among clinicians.”

Professor Namkug Kim of the Department of Convergence Medicine at Asan Medical Center said, “We are focusing on developing technologies that are essential in clinical settings but have not yet been sufficiently studied, such as neonatal intestinal perforation. We aim to create and apply various models that support early diagnosis in neonatal intensive care units, where rapid decision-making is critical, ultimately contributing to improving newborn survival rates.”

The results of this study were published in the latest issue of Computers in Biology and Medicine (impact factor 6.3), a leading international journal in the field of biomedical science.