GliomaGen: Conditional Diffusion for Post-Treatment Glioma MRI Generation
GliomaGen is a generative diffusion model tailored for synthesizing post-treatment glioma MRI images based on anatomical masks. It leverages a modified Med-DDPM architecture to create high-fidelity MRI images conditioned on segmented anatomical features.
Model Overview
GliomaGen aims to address data scarcity in post-treatment glioma segmentation tasks by expanding existing datasets with synthetic, high-quality MRI volumes. The model takes anatomical masks as input and generates multi-modal MRI scans conditioned on segmentation labels.
Model Performance
Quantitative Metrics
| Modality | FID (↓) | KID (↓) | MS-SSIM (↑) |
|---|---|---|---|
| t1c | 55.20 ± 3.74 | 0.0293 ± 0.0019 | 0.7647 ± 0.2106 |
| t2w | 54.99 ± 3.23 | 0.0291 ± 0.0010 | 0.6513 ± 0.2881 |
| t1n | 58.46 ± 3.86 | 0.0305 ± 0.0011 | 0.7005 ± 0.2585 |
| t2f | 70.42 ± 4.17 | 0.0370 ± 0.0018 | 0.7842 ± 0.1551 |
Usage
To use GliomaGen for MRI generation, see the GitHub repository.
BraTS 2024 Adult Post-Treatment Glioma-Synthetic
Alongisde GliomaGen, a synthetic dataset of $N=2124$ MR images is released on HuggingFace.
Inference Providers
NEW
This model is not currently available via any of the supported Inference Providers.
The model cannot be deployed to the HF Inference API:
The model has no library tag.
Evaluation results
- FID (t1c) on BraTS2024 Adult Post-Treatment Gliomaself-reported55.2028 ± 3.7446
- FID (t2w) on BraTS2024 Adult Post-Treatment Gliomaself-reported54.9974 ± 3.2271
- KID (t1c) on BraTS2024 Adult Post-Treatment Gliomaself-reported0.0293 ± 0.0019
- MS-SSIM (t1c) on BraTS2024 Adult Post-Treatment Gliomaself-reported0.7647 ± 0.2106