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

Ha Kyung Jung^1,†   ·   Changyong Choi^2,3,†   ·   Ji Eun Park^4,*   ·   Seo Young Park⁵   ·   Jae Ho Lee⁶   ·   Namkug Kim^2,4   ·   Ho Sung Kim⁴

¹Department of Radiology, Keimyung University Dongsan Hospital, Keimyung University School of Medicine, Daegu, Korea
²Department of Convergence Medicine, Asan Medical Center, University of Ulsan College of Medicine
³Department of Biomedical Engineering, AMIST, Asan Medical Center, University of Ulsan College of Medicine
⁴Department of Radiology and Research Institute of Radiology, Asan Medical Center, University of Ulsan College of Medicine, Seoul, Korea
⁵Department of Statistics and Data Science, Korea National Open University, Seoul, Korea
⁶Department of Korea and Center for Imaging Science, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea

† equal contribution
* corresponding author

Paper

Abstract

This study investigated the effects of feature augmentation, which uses generated images with specific imaging features, on the performance of isocitrate dehydrogenase (IDH) mutation prediction models in gliomas. A total of 598 patients were included from our institution (310 training, 152 internal test) and the Cancer Genome Atlas (136 external test). Score-based diffusion models were used to generate T2-weighted, FLAIR, and contrast-enhanced T1-weighted image triplets. Three neuroradiologists independently assessed visual Turing tests and various morphological features. Multivariable logistic regression models were developed using real images, random augmented data, and feature-augmented datasets. While random augmentation yielded models with AUCs comparable to real image-based models, it led to reduced specificity, particularly in the external test set (specificity: 83.2% vs. 73.0%, P = .013). In contrast, feature-augmented models maintained stable diagnostic performance; however, when more than 70% of training images included synthetic T2-FLAIR mismatch signs, AUC decreased in the external test set (AUC: 0.905–0.906 for ≤ 70%; 0.902–0.876 for ≥ 80%). These findings highlight the value of phenotype-specific augmentation for IDH prediction, while emphasizing the need to optimize augmentation proportion to avoid performance degradation.

Dependencies

Install the other packages in requirements.txt, jax, jaxlib, numpy, and opencv-python as following:

pip install -r requirements.txt
pip install jax==0.4.6 jaxlib==0.4.6 -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.htm
pip install numpy==1.23.0
pip install opencv-python==4.5.5.64

Prepare your own dataset

For example, you should set dataset path following:

root_path
    ├── train
          ├── <Patient_Folder>
                ├── T1CE
                      ├── 0001.npy
                      ├── 0002.npy
                      └── 0003.npy
                ├── T2
                └── FLAIR
    └── test

Training

python main.py --config='configs/ve/t1t2flair.py' --workdir='result' --mode=train

Model checkpoints and validation samples will be stored in ./result/checkpoints and ./result/samples, respectively.

Sampling

python t1t2flair_sampling.py

Sampling results will be stored in ./result/generated_images as png file.

Acknowledgement

Our main code is heavily based on score_sde_pytorch.

BibTeX

@article{jung2025idh,
  title={Phenotype augmentation using generative AI for isocitrate dehydrogenase mutation prediction in glioma},
  author={Jung, Ha Kyung and Choi, Changyong and Park, Ji Eun and Park, Seo Young and Lee, Jae Ho and Kim, Namkug and Kim, Ho Sung},
  journal={Scientific Reports},
  volume={15},
  number={1},
  pages={28913},
  year={2025},
  publisher={Nature Publishing Group UK London},
  doi={10.1038/s41598-025-14477-z}
}