DENOISING ON MONTE CARLO DOSE DISTRIBUTIONS

放射治疗计划需要精确的剂量计算，以便最大限度地发挥精确定向放射的杀肿瘤作用，同时最大限度地减少传递到附近正常组织的剂量。蒙特卡罗 (MC) 辐射传输是唯一能够在所有临床感兴趣的情况下满足这一需求的方法。然而，必须运行 MC 剂量计算，直到所得剂量分布中的统计波动（“噪声”）充分减少。我们发现，通过从噪声模拟输出中对实际无噪声基础剂量分布​​进行统计估计，可以大大提高 MC 精度，我们将这一过程称为“去噪”。我们已经证明，去噪能够将 MC 计算时间减少至少几倍。我们建议研究去噪在适形光子治疗和调强放射治疗（IMRT）剂量计算中的应用。量化降噪性能的指标将在具体目标#1 下制定。将在特定目标 #2 下开发 MC 剂量分布基准测试套件，包括光子束、IMRT 笔形束和优化的 IMRT 剂量分布。小波收缩阈值去噪将在特定目标#3 下开发。使用空间自适应迭代滤波的去噪将在特定目标#4下开发。两种降噪方法的相对性能和临床可接受性将根据基准测试套件以及根据特定目标 #1 开发的指标进行测试。在具体目标#5下，我们建议使用离散小波变换来降噪和压缩三维MC生成的笔形射束（PB）剂量分布，并有效计算IMRT注量加权PB剂量分布。具体目标#6 将确定 IMRT 治疗计划可接受的最大 MC PB 噪声水平。我们假设外部光子束和 IMRT PB 的最佳去噪算法将减少 MC 计算时间至少 5-10 倍。我们进一步假设基于小波的剂量计算方法将：（a）能够使用精确的基于 MC 的 PB 剂量分布进行 IMRT 治疗计划，（b）适用于 MC 或任何其他 PB 剂量计算算法，以及（c）在每次 IMRT 优化迭代中比完整剂量重新计算的计算效率更高。这些结果将实现我们提高放射治疗计划的临床有效性的总体目标。

Radiation therapy treatment planning requires accurate dose calculations in order to maximize the tumoricidal effects of precisely directed radiation while minimizing doses delivered to nearby normal tissues. Monte Carlo (MC) radiation transport is the only method which is capable of fulfilling this need in all situations of clinical interest. However, MC dose computations must be run until statistical fluctuations ("noise") in the resulting dose distributions are adequately reduced. We have discovered that MC precision can be greatly improved through statistical estimation of the actual noise-free underlying dose distribution from the noisy simulation output, a process we term "denoising." We have shown that denoising is capable of reducing MC calculation times at least several-fold. We propose to investigate the application of denoising to conformal photon therapy and intensity modulated radiation therapy (IMRT) dose calculations. Metrics for quantifying denoising performance will be developed under Specific Aim #1. A benchmark test suite of MC dose distributions, including photon beam, IMRT pencil beam, and optimized IMRT dose distributions, will be developed under Specific Aim #2. Wavelet shrinkage threshold denoising will be developed under Specific Aim #3. Denoising using spatially adaptive iterative filtering will be developed under Specific Aim #4. The relative performance and clinical acceptability of the two denoising methods will be tested against the benchmark test suite with the metrics developed under Specific Aim #1. Under Specific Aim #5 we propose to use discrete wavelet transforms to denoise and compress three dimensional MC- generated pencil beam (PB) dose distributions, and to efficiently compute IMRT fluence-weighted PB dose distributions. Specific Aim #6 will establish maximum MC PB noise levels acceptable for IMRT treatment planning. We hypothesize that optimal denoising algorithms for external photon beams and IMRT PBs will decrease MC computation times by at least a factor of 5-10. We further hypothesize that wavelet-based dose computation methods will: (a) enable use of accurate MC-based PB dose distributions for IMRT treatment planning, (b) apply to MC or any other PB dose calculation algorithm, and (c) be far more computationally efficient than complete dose recalculations at each IMRT optimization iteration. These results would achieve our overall goal of increasing the clinical effectiveness of radiation therapy treatment planning.