TLDR: A groundbreaking study by researchers at Asan Medical Center has demonstrated that generative artificial intelligence (AI) can significantly improve the accuracy of predicting isocitrate dehydrogenase (IDH) mutations in glioma, a critical biomarker for brain tumor prognosis. The AI-powered method, utilizing phenotype augmentation, outperformed experienced neuroradiologists in diagnostic precision, offering a promising new avenue for non-invasive and efficient glioma diagnosis.
Seoul, South Korea – In a significant advancement for neuro-oncology, a research team from Asan Medical Center has successfully leveraged generative artificial intelligence (AI) to enhance the prediction of isocitrate dehydrogenase (IDH) mutations in glioma, a type of brain tumor. The study, led by Professors Hye Hyeon Moon and Ji Eun Park from the Department of Radiology, and Professor Namkug Kim from the Department of Convergence Medicine, highlights the potential of AI-based augmentation to overcome limitations in deep learning models caused by insufficient real-world patient data.
IDH mutations are recognized as crucial molecular biomarkers that profoundly influence the prognosis and treatment strategies for glioma patients. Accurate and timely diagnosis of these mutations is therefore paramount for effective patient management. Traditionally, diagnosing genetic mutations using deep learning has faced challenges due to the scarcity of diverse and sufficient real-world patient data for training robust models.
The researchers developed a Generative AI-based Augmentation (GAA) system, employing a score-based diffusion model and a ResNet50 classifier. This system was trained using contrast-enhanced T1 and FLAIR MRI images from 565 glioma patients, comprising 346 IDH-wildtype and 219 IDH-mutant cases, collected between August 2008 and September 2020. The GAA was capable of generating realistic and diverse imaging phenotypes of glioma, effectively augmenting the training dataset.
The study found that the IDH mutation prediction accuracy was optimized when the number of generated images reached 110,000. The performance of the AI model was rigorously tested through both internal and external validation sets. The internal validation included data from 119 patients (78 IDH-wildtype, 41 IDH-mutant), while the external validation used data from 108 patients (72 IDH-wildtype, 36 IDH-mutant).
Remarkably, the AI model achieved an impressive accuracy rate of 93.8% in the internal validation and 83.3% in the external validation. These figures significantly surpassed the diagnostic accuracy of human neuroradiologists. For comparison, neuroradiologists with eight years of experience achieved 86.4% accuracy internally and 81.9% externally, while those with two years of experience showed 79% internal and 74.3% external accuracy. The AI classifier demonstrated statistically higher Area Under the Receiver Operating Characteristics Curve (AUC) values compared to neuroradiologists across both test sets (internal: P = .003 and P < .001; external: P < .01 and P < .001).
The generated images themselves demonstrated high realism, passing a Turing’s test with 47.5–50.5% success, indicating their indistinguishability from real MRI scans. This realism and diversity are key to the AI’s enhanced diagnostic capability.
Also Read:
- Google’s Med-Gemini AI Identifies Non-Existent Brain Structure, Raising Medical Safety Concerns
- Rowan University AI Innovator Honored with Prestigious 2025 Edison Patent Award for Cancer Diagnostics
The findings of this pivotal research were published in the leading neuro-oncology journal, ‘Neuro-Oncology,’ in June 2024. This study marks a significant step forward in integrating advanced AI technologies into clinical diagnostics, promising more accurate and efficient non-invasive prediction of IDH mutations in glioma, ultimately aiding in better patient outcomes.
