Department of Nuclear and Radiological Engineering, Georgia institute of Technology, Atlanta, GA, USA.
Department of Radiation Oncology, Emory University, Atlanta, GA, USA.
Acta Oncol. 2023 Jun;62(6):627-634. doi: 10.1080/0284186X.2023.2224050. Epub 2023 Jun 19.
Because proton head and neck (HN) treatments are sensitive to anatomical changes, plan adaptation (re-plan) during the treatment course is needed for a significant portion of patients. We aim to predict re-plan at plan review stage for HN proton therapy with a neural network (NN) model trained with patients' dosimetric and clinical features. The model can serve as a valuable tool for planners to assess the probability of needing to revise the current plan.
Mean beam dose heterogeneity index (BHI), defined as the ratio of the maximum beam dose to the prescription dose, plan robustness features (clinical target volume (CTV), V100 changes, and V100 > 95% passing rates in 21 robust evaluation scenarios), as well as clinical features (e.g., age, tumor site, and surgery/chemotherapy status) were gathered from 171 patients treated at our proton center in 2020, with a median age of 64 and stages from I-IVc across 13 HN sites. Statistical analyses of dosimetric parameters and clinical features were conducted between re-plan and no-replan groups. A NN was trained and tested using these features. Receiver operating characteristic (ROC) analysis was conducted to evaluate the performance of the prediction model. A sensitivity analysis was done to determine feature importance.
Mean BHI in the re-plan group was significantly higher than the no-replan group ( < .01). Tumor site ( < .01), chemotherapy status ( < .01), and surgery status ( < .01) were significantly correlated to re-plan. The model had sensitivities/specificities of 75.0%/77.4%, respectively, and an area under the ROC curve of .855.
There are several dosimetric and clinical features that correlate to re-plans, and NNs trained with these features can be used to predict HN re-plans, which can be used to reduce re-plan rate by improving plan quality.
由于质子头颈部（HN）治疗对解剖结构变化敏感，因此很大一部分患者在治疗过程中需要进行计划调整（重新计划）。我们旨在通过使用患者剂量学和临床特征训练的神经网络（NN）模型在计划审查阶段预测 HN 质子治疗的重新计划。该模型可以作为规划师评估需要修改当前计划的可能性的有价值的工具。
从 2020 年在我们质子中心接受治疗的 171 名患者中收集了平均射束剂量不均匀性指数（BHI）、计划稳健性特征（CTV、V100 变化以及在 21 个稳健性评估场景中 V100＞95%通过率）以及临床特征（例如年龄、肿瘤部位和手术/化疗状态）。对重新计划和无需重新计划组的剂量学参数和临床特征进行了统计分析。使用这些特征对 NN 进行了训练和测试。进行了接收者操作特征（ROC）分析以评估预测模型的性能。进行了敏感性分析以确定特征的重要性。
重新计划组的平均 BHI 明显高于无需重新计划组（ <.01）。肿瘤部位（ <.01）、化疗状态（ <.01）和手术状态（ <.01）与重新计划明显相关。该模型的灵敏度/特异性分别为 75.0%/77.4%，ROC 曲线下面积为.855。
有几个剂量学和临床特征与重新计划相关，并且使用这些特征训练的 NN 可用于预测 HN 重新计划，从而通过提高计划质量来降低重新计划率。
