| 陈路桥,倪千喜,吴宇,等.基于放射组学特征预测γ通过率的多中心研究[J].中华放射医学与防护杂志,2024,44(12):1027-1033.Chen Luqiao,Ni Qianxi,Wu Yu,et al.A multicenter study on the prediction of gamma passing rate based on radiomic features[J].Chin J Radiol Med Prot,2024,44(12):1027-1033 |
| 基于放射组学特征预测γ通过率的多中心研究 |
| A multicenter study on the prediction of gamma passing rate based on radiomic features |
| 投稿时间：2023-12-11 |
| DOI：10.3760/cma.j.cn112271-20231211-00205 |
| 中文关键词: 机器学习 容积旋转调强放疗 放射组学 γ通过率 多中心 |
| 英文关键词:Machine learning Volumetric-modulated arc therapy Radiomics Gamma passing rate Multicenter |
| 基金项目:湖南省自然科学基金(2023JJ30373);湖南省科技创新计划资助项目(2021SK51116,2023SK4034);湖南省卫生健康委适宜技术推广项目(202218015767);湖南省肿瘤医院攀登科研计划重点研发项目(YF2021006) |
|
| 摘要点击次数: 1474 |
| 全文下载次数: 559 |
| 中文摘要: |
| 目的 采用基于放射组学的机器学习方法,在多个放疗机构中构建γ通过率分类预测模型,并评估模型的分类性能。方法 回顾性收集了来自3个放疗机构的572例容积旋转调强放疗(VMAT)患者数据(其中514例作为训练集,58例用作测试集),额外收集了单个机构的45例VMAT计划作为独立的外部验证集,均使用基于模体实际测量的三维剂量验证方式并在3%/2 mm标准下采用10%剂量阈值、绝对剂量及全局归一进行γ分析。提取基于剂量文件的放射组学特征,使用随机森林(RF)方法以及RF结合沙普利加性解释(SHAP)方法进行特征筛选,根据特征排序分别选择不同数量(10、20、30、40、50)的特征子集作为模型的输入,使用极端梯度提升树算法对数据进行训练,通过受试者工作特征曲线下面积(AUC)值及F1分数来评估模型分类性能。结果 在3%/2 mm标准下,均在特征子集数量为20时模型性能表现最佳。在测试集和外部验证集中,经RF特征选择的最佳预测模型的AUC值和F1分数分别为0.88和0.89、0.82和0.90;经RF-SHAP特征选择的最佳预测模型的AUC值和F1分数分别为0.86和0.92、0.87和0.89,且经RF-SHAP特征选择的最佳模型表现出更好的稳健性,与RF特征选择方法相比具有一定的优势。结论 针对多中心的剂量验证结果,可以使用基于剂量文件的放射组学特征结合基于SHAP值的特征选择方法来构建机器学习预测模型并具有较好的分类性能,有助于推进伽马通过率预测模型的临床应用与实施。 |
| 英文摘要: |
| Objective To construct classification prediction models for gamma passing rate using radiomics-based machine learning approaches and data from multiple radiotherapy institutions and evaluate the models' performance. Methods The data from 572 volumetric-modulated arc therapy (VMAT) patients across three radiotherapy institutions (514 for training and 58 for testing) were retrospectively collected. Additionally, 45 VMAT plans were collected from a single institution as an independent external validation set. For all the data, a three-dimensional dose validation approach based on actual measurements of phantoms was utilized, and gamma analysis was performed at the 3%/2 mm criterion using a dose threshold of 10%, absolute doses, and global normalization. After radiomic features were extracted from dose files, feature selection was performed using the random forest (RF) method and RF combined with Shapley Additive exPlanation (SHAP). Then, feature subsets of varying sizes (10, 20, 30, 40, and 50) were selected based on feature rankings. Using these subsets as inputs, data training was conducted using the Extreme Gradient Boosting (XGBoost) algorithm. Finally, the models' classification performance was assessed using the area under the curve (AUC) values and F1-score. Results Under the 3%/2 mm criterion, all models performed the best in the case of 20 feature subsets. The optimal prediction model established based on the feature selection using RF exhibited AUC and F1-score of 0.88 and 0.89, respectively on the testing set and 0.82 and 0.90, respectively, on the validation set. The optimal prediction model built based on the feature selection using RF combined with SHAP yielded AUC and F1-score of 0.86 and 0.92 on the testing set and 0.87 and 0.89, respectively, on the validation set, along with superior robustness. Therefore, the second model possessed certain advantages over the first model. Conclusions For multicenter dose verification result, it is feasible to construct a machine learning prediction model with high classification performance using radiomic features derived from dose files, combined with feature selection based on SHAP. This approach can assist in advancing the clinical applications and implementation of gamma passing rate prediction models. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|