Diffusion Transformer-based Universal Dose Denoising for Pencil Beam Scanning Proton Therapy.

Purpose: Intensity-modulated proton therapy (IMPT) is an advanced treatment modality for head and neck (H&N) cancer patients, offering precise tumor dose coverage while sparing surrounding organs at risk (OARs). However, IMPT is highly sensitive to inter-fractional anatomical changes, necessitating periodic adjustments through online adaptive radiation therapy (oART). But a significant bottleneck in the current oART workflow is the need for fast and accurate dose calculation using Monte Carlo (MC) simulations for plan quality assessment and re-optimization. Reducing the number of particles in MC-based simulations can accelerate dose calculation but at the cost of reduced accuracy. To address this, denoising noisy dose maps generated by low statistics MC simulations has been proposed as a method to rapidly and accurately generate high-accuracy dose maps. Methods: A diffusion transformer-based dose denoising framework was developed. IMPT treatment plans and 3D CT images from 80 H&N cancer patients were used to construct the training dataset by generating noisy dose maps and their corresponding high statistics dose maps using an open-source MC software, MCsquare, with a computation time of approximately 1 minutes and 10 minutes per plan, respectively. Each data sample was standardized into uniform chunks with zero-padding. Then, normalization and non-linear mapping were applied to the data samples to transform them toward a quasi-Gaussian distribution. The treatment plans and 3D CT images from another independent 10 H&N cancer patients, 10 prostate cancer patients, 10 lung cancer patients, and 10 breast cancer patients were used as the testing dataset, following the same preprocessing protocol as the training dataset. The proposed model was trained with noisy dose maps and 3D CT images as input and high statistics dose maps as the ground truth. The training was constrained by mean square error (MSE) loss, a residual loss that focused on reducing the difference between the predicted and ground truth dose maps and a regional mean absolute error (MAE) loss that specifically targeted voxels with the top 10% and bottom 10% dose values. Performance was evaluated using MAE. 3D Gamma passing rates and dose volume histogram (DVH) indices were calculated to assess differences between the predicted and ground-truth dose maps. Results: The proposed framework achieved MAE of 0.195 ± 0.112 Gy[RBE], 0.120 ± .054 Gy[RBE], 0 . 172 ± .096 Gy[RBE], and 0.376 ± 0.375 Gy[RBE] for H&N, lung, breast and prostate testing cases, respectively. The 3D gamma passing rate consistently exceeded 92% in the whole body using a 3%/2 mm criterion across all disease sites. DVH indices calculated from the ground truth and predicted dose distributions showed excellent agreement for both clinical target volumes (CTVs) and OARs. Conclusion: A diffusion transformer-based denoising framework was successfully developed. Although the denoising model was trained using only H&N data, it can accurately and robustly denoise noisy dose maps across different disease sites.