A South Korean research team has developed an artificial intelligence (AI) system that can predict tumor genetic mutations and automatically generate radiology reports from a single brain MRI scan, eliminating the need for a biopsy.
The team — led by Professor Park Sang-hyun of the Department of Computer Science and Engineering at Pohang University of Science and Technology (POSTECH), master's student Ryu Hee-seung of POSTECH's Graduate School of Artificial Intelligence, Dr. Kang Myung-kyun of the Daegu Gyeongbuk Institute of Science and Technology (DGIST), and Professors Park Ye-won and Ahn Sung-soo of Severance Hospital — developed a vision-language AI model that predicts brain tumor characteristics from MRI images and automatically generates radiology reports.
Adult-type diffuse glioma, a form of malignant brain tumor, is characterized by significant variation in treatment options and prognosis from patient to patient. As a result, the presence of mutations in the isocitrate dehydrogenase (IDH) gene — a normal protein used by cells in the body for energy production — is used as critical information in determining the treatment direction for each patient.
Previously, confirming this required directly extracting brain tissue, placing a heavy burden on patients and posing limitations in testing time and cost. While research has recently been actively conducted to non-invasively predict patient status using MRI, the increased volume of MRI scans has heavily burdened medical staff with reading and reporting duties.
The research team developed a new AI model, "Glio-LLaMA-Vision," that can address both of these issues simultaneously.
The AI model successfully predicts IDH mutation status from brain MRI images alone. It demonstrated AUC performance of 0.85 to 0.95 (0.9 or higher = very high performance) in prediction accuracy and operated stably across various environments.
The AI-generated radiology reports also received high marks from specialists. More than approximately 90% were judged suitable for use in actual clinical practice, with some evaluated as equal to or better than existing radiology reports.
The team explained that to achieve such strong performance, they trained the AI on large-scale biomedical image-text data using PubMed Central (PMC), a biomedical research paper database. They then further trained the model on MRI images and actual radiology report data from brain tumor patients to develop it into a model specialized in brain tumors.
In this process, they also applied a preprocessing step to organize and standardize sentence formats to reduce inconsistencies in radiology report expressions. This more stable training enabled the AI to generate radiology reports in a consistent format.
The research is significant in that it links MRI analysis, genetic prediction, and radiology report writing into a single AI system. It demonstrates the potential to assist the entire diagnostic process, going beyond simply aiding analysis.
"We confirmed the potential to reduce the burden of image reading while helping enable rapid treatment decisions without genetic testing," Professor Park Sang-hyun of POSTECH said. "Through follow-up research, we will develop this into a medical AI that can be used in actual clinical practice."
The research was conducted with funding from the Ministry of Science and ICT and support from the National Research Foundation of Korea, and was published in the international healthcare journal "npj Digital Medicine."
