3D-GlioPREDICT uses latent diffusion to predict post-radiotherapy brain MRI in glioma patients
A 3D AI model synthesizes post-treatment brain scans from pretreatment data and dose maps
Radiotherapy is a cornerstone of glioma treatment, but it induces complex structural brain changes that are difficult to anticipate. Predicting these changes from pretreatment data could improve understanding of treatment effects and support outcome prediction. Previous methods mostly operated on single 2D slices and treated radiation as a global parameter. Now, a team from the University Medical Center Utrecht introduces 3D-GlioPREDICT, a 3D latent diffusion framework that conditions image generation on spatially resolved voxel-wise dose distributions, alongside pretreatment MRI and follow-up time. This makes volumetric synthesis computationally feasible by combining latent-space compression with ControlNet-based spatial conditioning.
Trained on a public dataset of 257 scans from 25 glioma patients, 3D-GlioPREDICT was evaluated using mean squared error, peak signal-to-noise ratio, and structural similarity index. Anatomical consistency was further assessed via Dice scores for cerebrospinal fluid, gray matter, and white matter segmentations, plus hippocampus volume prediction error and deformation analysis. Compared to the team’s earlier 2D model, the 3D version achieved better image similarity while maintaining high agreement with ground truth anatomy and deformation patterns. These results demonstrate the feasibility of 3D treatment-aware generative modeling for predicting post-radiotherapy brain MRI from only pretreatment information. Code is released on GitHub.
- 3D-GlioPREDICT uses a 3D latent diffusion model conditioned on voxel-wise radiation dose and follow-up time, unlike prior 2D slice-based methods that treat dose as a global parameter.
- The model was trained and evaluated on 257 scans from 25 glioma patients, outperforming 2D baselines on image similarity metrics (MSE, PSNR, SSIM) and anatomical Dice scores for CSF, gray matter, and white matter.
- Code is publicly available on GitHub, enabling reproducibility and further research into AI-driven treatment effect prediction.
Why It Matters
Enables personalized radiotherapy planning by predicting structural brain changes, potentially improving outcome forecasting and adaptive treatment.