Shah Keyur D, Yeap Beow Y, Lee Hoyeon, Soetan Zainab O, Moteabbed Maryam, Muise Stacey, Cowan Jessica, Remillard Kyla, Silvia Brenda, Mendenhall Nancy P, Soffen Edward, Mishra Mark V, Kamran Sophia C, Miyamoto David T, Paganetti Harald, Efstathiou Jason A, Chamseddine Ibrahim
Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA.
Department of Radiation Oncology, Winship Cancer Institute, Emory University, Atlanta, GA.
JCO Clin Cancer Inform. 2025 Mar;9:e2400252. doi: 10.1200/CCI-24-00252. Epub 2025 Mar 19.
To aid personalized treatment selection, we developed a predictive model for acute rectal toxicity in patients with prostate cancer undergoing radiotherapy with photons and protons.
We analyzed a prospective multi-institutional cohort of 278 patients treated from 2012 to 2023 across 10 centers. Dosimetric and nondosimetric variables were collected, and key predictors were identified using purposeful feature selection. The cohort was split into discovery (n = 227) and validation (n = 51) data sets. The dose along the rectum surface was transformed into a two-dimensional surface, and dose-area histograms (DAHs) were quantified. A convolutional neural network (CNN) was developed to extract dosimetric features from the DAH and integrate them with nondosimetric predictors. Model performance was benchmarked against logistic regression (LR) using the AUC.
Key predictors included rectum length, race, age, and hydrogel spacer use. The CNN model demonstrated stability in the discovery data set (AUC = 0.81 ± 0.11) and outperformed LR in the validation data set (AUC = 0.81 0.54). Separate analysis of photon and proton subsets yielded consistent AUCs of 0.7 and 0.92, respectively. In the photon high-risk group, the model achieved 83% sensitivity, and in proton subsets, it achieved 100% sensitivity and specificity, indicating the potential to be used for treatment selection in these patients.
Our novel approach effectively predicts rectal toxicity across photon and proton subsets, demonstrating the utility of integrating dosimetric and nondosimetric features. The model's strong performance across modalities suggests potential for guiding treatment decisions, warranting prospective validation.
为辅助个性化治疗选择，我们针对接受光子和质子放疗的前列腺癌患者开发了一种急性直肠毒性预测模型。
我们分析了2012年至2023年期间在10个中心接受治疗的278例患者的前瞻性多机构队列。收集剂量学和非剂量学变量，并使用有目的的特征选择确定关键预测因素。该队列被分为发现数据集（n = 227）和验证数据集（n = 51）。将直肠表面的剂量转换为二维表面，并对剂量-面积直方图（DAH）进行量化。开发了一种卷积神经网络（CNN），以从DAH中提取剂量学特征，并将其与非剂量学预测因素整合。使用AUC将模型性能与逻辑回归（LR）进行基准比较。
关键预测因素包括直肠长度、种族、年龄和水凝胶间隔物的使用。CNN模型在发现数据集中表现出稳定性（AUC = 0.81 ± 0.11），在验证数据集中优于LR（AUC = 0.81 0.54）。对光子和质子亚组的单独分析分别产生了一致的AUC，分别为0.7和0.92。在光子高风险组中，该模型的灵敏度达到83%，在质子亚组中，其灵敏度和特异性均达到100%，表明该模型有潜力用于这些患者的治疗选择。
我们的新方法有效地预测了光子和质子亚组中的直肠毒性，证明了整合剂量学和非剂量学特征的实用性。该模型在不同模式下的强大性能表明其在指导治疗决策方面具有潜力，值得进行前瞻性验证。
