Department of Radiation Oncology, Yonsei Cancer Center, Seoul, Republic of Korea.
Department of Radiation Oncology, Asan Medical Center, Seoul, Republic of Korea.
Med Phys. 2023 Nov;50(11):7203-7213. doi: 10.1002/mp.16646. Epub 2023 Jul 30.
Many studies have utilized optical camera systems with volumetric scintillators for quality assurances (QA) to estimate the proton beam range. However, previous analytically driven range estimation methods have the difficulty to derive the dose distributions from the scintillation images with quenching and optical effects.
In this study, a deep learning method utilized to QA was used to predict the beam range and spread-out Bragg peak (SOBP) for two-dimensional (2D) map conversion from the scintillation light distribution (LD) into the dose distribution in a water phantom.
The 2D residual U-net modeling for deep learning was used to predict the 2D water dose map from a 2D scintillation LD map. Monte Carlo simulations for dataset preparation were performed with varying monoenergetic proton beam energies, field sizes, and beam axis shifts. The LD was reconstructed using photons backpropagated from the aperture as a virtual lens. The SOBP samples were constructed based on monoenergetic dose distributions. The training set, including the validation set, consisted of 8659 image pairs of LD and water dose maps. After training, dose map prediction was performed using a 300 image pair test set generated under random conditions. The pairs of simulated and predicted dose maps were analyzed by Bragg peak fitting and gamma index maps to evaluate the model prediction.
The estimated beam range and SOBP width resolutions were 0.02 and 0.19 mm respectively for varying beam conditions, and the beam range and SOBP width deviations from the reference simulation result were less than 0.1 and 0.8 mm respectively. The simulated and predicted distributions showed good agreement in the gamma analysis, except for rare cases with failed gamma indices in the proximal and field-marginal regions.
The deep learning conversion method using scintillation LDs in an optical camera system with a scintillator is feasible for estimating proton beam range and SOBP width with high accuracy.
许多研究都利用带有体闪烁器的光学相机系统进行质量保证 (QA) 来估计质子束射程。然而，之前基于分析的射程估计方法在从具有淬灭和光学效应的闪烁图像推导剂量分布方面存在困难。
本研究利用用于 QA 的深度学习方法，将闪烁光分布 (LD) 转换为水模中的剂量分布，从二维 (2D) 闪烁 LD 图预测二维 (2D) 地图的束射程和扩展布拉格峰 (SOBP)。
使用二维残差 U-net 建模进行深度学习，从 2D 闪烁 LD 图预测 2D 水剂量图。为数据集准备进行了蒙特卡罗模拟，模拟了不同单能质子束能量、射野大小和束轴偏移。使用从孔径反向传播的光子作为虚拟透镜重建 LD。基于单能剂量分布构建 SOBP 样本。训练集包括验证集，由 8659 对 LD 和水剂量图的图像对组成。训练后，使用在随机条件下生成的 300 对测试图像对进行剂量图预测。通过 Bragg 峰拟合和伽马指数图分析模拟和预测的剂量图对来评估模型预测。
对于不同的束条件，估计的束射程和 SOBP 宽度分辨率分别为 0.02 和 0.19 mm，束射程和 SOBP 宽度与参考模拟结果的偏差分别小于 0.1 和 0.8 mm。除了在近场和场边区域极少数伽马指数失败的情况下，模拟和预测分布在伽马分析中显示出很好的一致性。
使用带有闪烁体的光学相机系统中的闪烁 LD 的深度学习转换方法可用于高精度地估计质子束射程和 SOBP 宽度。
