Rectified flow-based prediction of post-treatment brain MRI from pre-radiotherapy priors for patients with glioma
New rectified flow model generates realistic follow-up MRI scans in real-time, achieving 0.91 Dice score accuracy.
A research team from UMC Utrecht has developed a novel AI system that can predict what a patient's brain will look like after radiotherapy treatment. Using a rectified flow model conditioned on pre-treatment MRI scans and radiation dose maps, the system generates realistic follow-up MRI images for any time point during treatment. The model incorporates cross-attention mechanisms to integrate temporal data and chemotherapy information, creating comprehensive simulations of post-treatment morphological changes.
Validation results show impressive performance: the generated images achieve a structural similarity index (SSIM) of 0.88 and peak signal-to-noise ratio (PSNR) of 22.82 when compared to real follow-up scans. Most significantly, tissue segmentations from predicted versus real MRI yield a mean Dice-Sørensen coefficient of 0.91, indicating high anatomical accuracy. The rectified flow architecture enables up to 250x faster inference than traditional Denoising Diffusion Probabilistic Models (DDPMs), making real-time clinical applications feasible.
This technology enables counterfactual simulations where doctors can vary treatment parameters to predict different outcomes before administering actual radiation. By modeling how different dose distributions affect brain tissue over time, clinicians could optimize treatment plans for individual patients. The system uses the public SAILOR dataset of 25 glioma patients and represents a significant advancement in personalized oncology, potentially reducing the 20 years of life lost on average from brain tumors through more precise, adaptive treatment approaches.
- Generates realistic follow-up MRI with 0.91 Dice score accuracy for tissue segmentation
- Runs 250x faster than diffusion models, enabling real-time clinical applications
- Allows counterfactual simulations by varying treatment parameters before actual radiation
Why It Matters
Enables personalized radiation dose planning and outcome prediction, potentially improving survival rates for brain tumor patients.