A Novel Proton Therapy Treatment Planning Solution: A Transformer-Based Poisson Flow Generative Model for MR-Only Synthetic CT Generation
Abstract
Purpose
Precise proton therapy treatment planning relies on the synergy between the superior soft-tissue definition obtained from MRI and the accurate relative stopping power (RSP) maps derived from CT. However, standard workflows that utilize separate scans are often compromised by inherent image co-registration errors and the additional radiation burden of CT scans. To overcome these limitations, we propose generating synthetic CT images directly from MRI, ensuring perfect anatomical alignment between soft-tissue and dosimetric data while eliminating imaging radiation dose. Current methods based on diffusion models often “hallucinate” textures or blur bone interfaces, which leads to critical range uncertainties in complex anatomies. We propose a novel 2D Poisson Flow Generative Model (PFGM) ++ equipped with a Swin-FIR Transformer backbone to synthesize brain/ spine cases.
Methods
Model: Our SwinFIR-UNet architecture integrates Swin Transformer blocks to capture long-range anatomical dependencies with Spatial-Frequency Blocks (SFB) to preserve high-frequency details. Our diffusion model was trained using the PFGM++ framework. It projects datapoints from an (N+D) dimensional hyperplane (D=128) to the N-dimensional CT data manifold conditioned on corresponding input MRs. Evaluation: Generation quality was assessed using PSNR, SSIM, and Mean Absolute Error (MAE) compared to ground-truth planning CTs.
Results
Our Swin-FIR PFGM++ model achieved a Mean Absolute Error (MAE) of 43.29 ± 7.79 HU, an SSIM of 0.88 ± 0.03, and a Peak Signal-to-Noise Ratio (PSNR) of 27.3123 ± 2.00 dB b, for the brain/ spine cases (proton planning data like dose-volume histogram curves will be included if the abstract is accepted).
Conclusion
Our method offers a robust pathway toward MR-only proton therapy planning by ensuring optimal image quality and minimizing the risks of range uncertainty associated with conventional generative AI.