Department of Radiation Medical Science and Engineering, Faculty of Medicine, Hokkaido University, Sapporo 060-8638, Japan.
Phys Med Biol. 2020 Oct 26;65(21):215007. doi: 10.1088/1361-6560/aba164.
This study proposes a near-real-time spot-scanning proton dose calculation method with probabilistic uncertainty estimation using a three-dimensional convolutional neural network (3D-CNN). CT images and clinical target volume contours of 215 head and neck cancer patients were collected from a public database. 1484 and 488 plans were extracted for training and testing the 3D-CNN model, respectively. Spot beam data and single-field uniform dose (SFUD) labels were calculated for each plan using an open-source dose calculation toolkit. Variable spot data were converted into a fixed-size volume hereby called a 'peak map' (PM). 300 epochs of end-to-end training was implemented using sets of stopping power ratio and PM as input. Moreover, transfer learning techniques were used to adjust the trained model to SFUD doses calculated with different beam parameters and calculation algorithm using only 7.95% of training data used for the base model. Finally, accuracy of the 3D-CNN-calculated doses and model uncertainty was reviewed with several evaluation metrics. The 3D-CNN model calculates 3D proton dose distributions accurately with a mean absolute error of 0.778 cGyE. The predicted uncertainty is correlated with dose errors at high contrast edges. Averaged Sørensen-Dice similarity coefficients between binarized outputs and ground truths are mostly above 80%. Once the 3D-CNN model was well-trained, it can be efficiently fine-tuned for different proton doses by transfer learning techniques. Inference time for calculating one dose distribution is around 0.8 s for a plan using 1500 spot beams with a consumer grade GPU. A novel spot-scanning proton dose calculation method using 3D-CNN was developed. The 3D-CNN model is able to calculate 3D doses and uncertainty with any SFUD spot data and beam irradiation angles. Our proposed method should be readily extendable to other setups and plans and be useful for dose verification, image-guided proton therapy, or other applications.
本研究提出了一种基于三维卷积神经网络（3D-CNN）的近实时点扫描质子剂量计算方法，具有概率不确定性估计。从公共数据库中收集了 215 例头颈部癌症患者的 CT 图像和临床靶区轮廓。分别提取了 1484 个和 488 个计划用于训练和测试 3D-CNN 模型。使用开源剂量计算工具包为每个计划计算点束数据和单野均匀剂量（SFUD）标签。将可变点数据转换为固定大小的体积，称为“峰图”（PM）。使用停止功率比和 PM 作为输入，进行了 300 个回合的端到端训练。此外，使用迁移学习技术，仅使用基础模型训练数据的 7.95%，调整训练好的模型，使其适用于不同束参数和计算算法计算的 SFUD 剂量。最后，使用多种评估指标审查了 3D-CNN 计算剂量和模型不确定性的准确性。3D-CNN 模型能够准确计算 3D 质子剂量分布，平均绝对误差为 0.778 cGyE。预测的不确定性与高对比度边缘的剂量误差相关。二值化输出与地面真值之间的平均 Sørensen-Dice 相似系数大多高于 80%。一旦 3D-CNN 模型得到很好的训练，通过迁移学习技术，它可以有效地为不同的质子剂量进行微调。使用消费级 GPU 计算一个使用 1500 个点束的计划的剂量分布的推断时间约为 0.8 秒。开发了一种新的基于 3D-CNN 的点扫描质子剂量计算方法。3D-CNN 模型能够使用任何 SFUD 点数据和束照射角度计算 3D 剂量和不确定性。我们提出的方法应该很容易扩展到其他设置和计划，并可用于剂量验证、图像引导质子治疗或其他应用。
