AI Enables Precise Diagnosis of Pediatric Fatty Liver... Only 1.45% Different From MRI

[Sportschosun reporter Jang Jong-ho] An artificial intelligence (AI) model has been developed that can measure the amount of fat in a child's liver at a level comparable to MRI.

As obesity has risen rapidly in recent years, the number of pediatric fatty liver patients in Korea has also continued to increase. According to the 2024 Korea National Health and Nutrition Examination Survey (KNHANES), the obesity prevalence rate among children aged 6 to 11 and adolescents aged 12 to 18 rose by 56.3% and 31.3%, respectively, over the past 10 years.

Among noninvasive tests used to quantify liver fat, MRI-PDFF is considered the most accurate method. However, its high cost and longer examination time than ultrasound make it difficult for young children to remain still for extended periods. In some cases, sedation may be required. As a result, its use in real-world clinical settings, including repeated follow-up tests, has been limited.

Against this backdrop, a research team led by Professor Choi Ga-young of the Department of Radiology at Korea University Ansan Hospital and Professor Ham, Sung Won of the Biomedical Research Center focused on 'raw ultrasound signals (RF data)' to develop a more efficient and accurate diagnostic technology for pediatric fatty liver.

Conventional ultrasound images, or B-mode images, are black-and-white displays created by converting ultrasound signals into a format that is easier for humans to view. Medical staff assess fatty liver by looking at differences in image brightness, which is a qualitative evaluation rather than a numerical one. As a result, findings can vary depending on the examiner's skill, and mild fatty liver may be missed.

By contrast, RF data are raw high-frequency signals before they are converted into images. They contain more tissue information, including the intensity and frequency changes of ultrasound reflected from inside the tissue. In this study, AI was used to analyze these raw high-frequency signals, making it possible to quantify even subtle changes in liver tissue that are difficult to distinguish on standard ultrasound images.

The research team collected MRI-PDFF and RF data from 40 children and adolescents suspected of having fatty liver. They also developed a range of AI models based on machine learning and deep learning, then compared each model's diagnostic accuracy against MRI-PDFF.

As a result, the AI model that used RF data along with blood test liver function values (ALT) and UGAP, an index that quantifies how much ultrasound weakens as it passes through tissue, showed the highest accuracy. The liver fat amount predicted by the model differed from the MRI-PDFF results by an average of only about 1.45%, indicating very high accuracy.

The team said the high accuracy came from AI's ability to analyze fine signal information inside tissues that is compressed during the process of converting RF data into conventional ultrasound images. They also said the AI model could serve as a practical alternative to MRI-PDFF for screening and monitoring pediatric fatty liver.

Professor Choi Ga-young said, "Pediatric fatty liver is a condition that requires not only accurate early diagnosis but also long-term management and repeated follow-up during growth after diagnosis." She added, "Until now, quantitative evaluation of pediatric fatty liver has not been easy because of difficulties in obtaining patient cooperation during examinations and the burden of MRI testing. By combining AI technology with ultrasound, which can be performed relatively easily without radiation exposure, we confirmed the possibility of more accurately quantifying liver fat in pediatric patients."

Meanwhile, the study was published in the international journal Scientific Reports and was selected for an oral presentation at IPR 2026, the world's largest international pediatric radiology conference, held in Boston in early June, underscoring the study's excellence.

Jang Jong-ho, bellho@sportschosun.com