Bridging Cone-Beam CT to Stopping Power Ratio Maps: A Brownian‑Bridge Approach to Support Robust Adaptive Proton Therapy
Abstract
Purpose
Adaptive radiation therapy (ART) for proton therapy relies on routine offline CT simulations for updated anatomy. To facilitate online proton ART, we present our high‑fidelity regularized Brownian Bridge (rBBrg) framework to perform deterministic cross-modality image translation of stopping power ratio (SPR) maps directly from cone-beam CTs (CBCTs).
Methods
rBBrg consisted of two distinct but mutually reinforcing mappings: (1) generation mapping where a Brownian Bridge diffusion model predicted SPR from CBCT and (2) reconstruction mapping where a conditional generative adversarial network (cGAN) reconstructed CBCT from predicted SPR providing cycle consistency and adversarial learning. Efficient, deterministic prediction was achieved via one-step sampling. Four-fold cross validation and augmentation were implemented. Conventional cGAN with perceptual loss was implemented for comparison. Matched pairs of planning CT and CBCT from 16 head-and-neck cancer patients were evaluated. To facilitate comparisons, CBCTs were deformably registered to CTs. A calibration phantom was scanned to establish Hounsfield look-up table for ground truth SPR. Predicted SPR and HU were evaluated via mean absolute error (MAE), peak-signal-to-noise ratio (PSNR) and structural similarity (SSIM) between real and predictions followed by Mann-Whitney U-test.
Results
For SPR within the external, rBBrg outperformed cGAN with MAE=0.038±0.003 vs 0.040±0.003 a.u. and SSIM=0.94±0.01 vs 0.93±0.01 a.u. (p0.05). Within air and soft tissue, predictions were not statistically different using rBBrg and cGAN, while bone predictions were improved with rBBRg (MAE=0.060±0.007 vs 0.068±0.008 a.u. (∆~13%, p<0.01)). Mapping synthetic SPR to HU, rBBrg obtained MAE=53±4 HU, outperforming cGAN MAE=56±5 HU (p<0.01). Qualitatively, despite strong dental artifacts on CBCT, rBBrg generated more realistic SPR maps and better preserved anatomical fidelity compared to cGAN.
Conclusion
The feasibility of implementing state-of-the-art rBBrg model to generate high-fidelity SPR to support online proton ART has been established. Future work will integrate metal artifact correction and comprehensive dosimetric validation.