P03.19.B TEXT2GLIOMA: AUGMENTATION OF GLIOMA MRI WITH SYNTHETIC DATA CONDITIONED ON CLINICAL INFORMATION

Abstract BACKGROUND Uncommon subtypes of glioma such as IDH-mutant astrocytomas are often under-represented in brain cancer datasets, leading to poor prediction performance (e.g., segmentation) on these cases. Tumor shape, size, and peritumoral edema vary widely with grade and genetic profile. Our objective was to generate high-quality, diverse synthetic FLAIR volumes conditioned on patient clinical notes. This synthetic dataset could augment real datasets and improve downstream models for automated volumetry (e.g., for surgical planning) or radiomics-based classifiers. MATERIAL AND METHODS The University of San Francisco-Preoperative Diffuse Glioma MRI (UCSF-PDGM; N=495) and University of Pennsylvania-Glioblastoma (UPENN-GBM; N=152) 3D FLAIR MRI datasets were used to train the model (Text2Glioma). The model consists of a variational autoencoder (VAE), which produced a latent embedding of FLAIR volumes, used as input to a latent diffusion U-Net model (LDM) with 1000 noising/denoising timesteps. The LDM was conditioned at each timestep with clinical text embedding vectors. Clinical text consisted of a single line of the format “diagnosis of <Astrocytoma/Glioblastoma/Oligodendroglioma>, <IDH-mutant/IDH-wildtype>, <male/female> aged <age in years>, WHO Grade <2/3/4>”. This information was extracted from the publicly available clinical information data in the dataset repositories. RESULTS The VAE model training after 100 epochs reached a minimum loss of 0.31. The LDM model training after 1000 epochs achieved a minimum mean-squared error loss of 0.681. Synthetic FLAIR images produced by text input visually resembled realistic FLAIR volumes with tumors present. Synthetic tumors were visually inspected for unrealistic morphology, though none were present. 100 WHO grade 2-3 tumor phenotypes were generated (40 with Astrocytoma, IDH-wildtype; 28 with Astrocytoma, IDH-mutant; 32 with Oligodendroglioma, IDH-mutant, 1p/19q co-deleted). Genders of simulated dataset were equal; ages uniformly ranged from 17-95, the range of ages in the real datasets. CONCLUSION With larger and more diverse training datasets, supplemented by realistic synthetic data, more robust and generalizable tumor predictions are possible across the full spectrum of glioma presentations. Future releases of this code will generate matching ground-truth tumor segmentations for radiomics or tumor segmentation models.