Phenotype augmentation using generative AI for isocitrate dehydrogenase mutation prediction in glioma

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Gliomas are primary brain tumors with varying molecular subtypes.
• Isocitrate dehydrogenase (IDH) mutations are key prognostic and predictive biomarkers in gliomas.
• Accurate prediction of IDH mutation status from medical imaging is crucial for clinical management.

Purpose of the Study:

• To evaluate the impact of feature augmentation using generated synthetic brain MRI data on the performance of IDH mutation prediction models in gliomas.
• To compare the effectiveness of random augmentation versus phenotype-specific feature augmentation.

Main Methods:

• Utilized score-based diffusion models to generate synthetic T2-weighted, FLAIR, and contrast-enhanced T1-weighted MRI image triplets.
• Developed multivariable logistic regression models using real images, randomly augmented data, and feature-augmented datasets.
• Assessed model performance using Area Under the Curve (AUC) and specificity on internal and external test sets.

Main Results:

• Random augmentation maintained comparable AUC to real image models but reduced specificity, especially in external datasets (83.2% vs. 73.0%).
• Feature-augmented models demonstrated stable diagnostic performance.
• Excessive synthetic data (over 70% T2-FLAIR mismatch signs) in training led to decreased AUC in the external test set (0.902-0.876).

Conclusions:

• Phenotype-specific feature augmentation is beneficial for improving IDH mutation prediction models in gliomas.
• Optimizing the proportion of synthetic data is essential to prevent performance degradation and ensure model reliability.