PDF(Pubmed)
Abstract：
BACKGROUND: Prediction of non-perfusion volume ratio (NPVR) is critical in selecting patients with uterine fibroids who will potentially benefit from ultrasound-guided high-intensity focused ultrasound (HIFU) treatment, as it reduces the risk of treatment failure. The purpose of this study is to construct an optimal model for predicting NPVR based on T2-weighted magnetic resonance imaging (T2MRI) radiomics features combined with clinical parameters by machine learning.
METHODS: This retrospective study was conducted among 223 patients diagnosed with uterine fibroids from two centers. The patients from one center were allocated to a training cohort (n = 122) and an internal test cohort (n = 46), and the data from the other center (n = 55) was used as an external test cohort. The least absolute shrinkage and selection operator (LASSO) algorithm was employed for feature selection in the training cohort. The support vector machine (SVM) was adopted to construct a radiomics model, a clinical model, and a radiomics-clinical model for NPVR prediction, respectively. The area under the curve (AUC) and the decision curve analysis (DCA) were performed to evaluate the predictive validity and the clinical usefulness of the model, respectively.
RESULTS: A total of 851 radiomic features were extracted from T2MRI, of which seven radiomics features were screened for NPVR prediction-related radiomics features. The radiomics-clinical model combining radiomics features and clinical parameters showed the best predictive performance in both the internal (AUC = 0.824, 95% CI 0.693-0.954) and external (AUC = 0.773, 95% CI 0.647-0.902) test cohorts, and the DCA also suggested the radiomics-clinical model had the highest net benefit.
CONCLUSIONS: The radiomics-clinical model could be applied to the NPVR prediction of patients with uterine fibroids treated by HIFU to provide an objective and effective method for selecting potential patients who would benefit from the treatment mostly.
METHODS: This retrospective study was conducted among 223 patients diagnosed with uterine fibroids from two centers. The patients from one center were allocated to a training cohort (n = 122) and an internal test cohort (n = 46), and the data from the other center (n = 55) was used as an external test cohort. The least absolute shrinkage and selection operator (LASSO) algorithm was employed for feature selection in the training cohort. The support vector machine (SVM) was adopted to construct a radiomics model, a clinical model, and a radiomics-clinical model for NPVR prediction, respectively. The area under the curve (AUC) and the decision curve analysis (DCA) were performed to evaluate the predictive validity and the clinical usefulness of the model, respectively.
RESULTS: A total of 851 radiomic features were extracted from T2MRI, of which seven radiomics features were screened for NPVR prediction-related radiomics features. The radiomics-clinical model combining radiomics features and clinical parameters showed the best predictive performance in both the internal (AUC = 0.824, 95% CI 0.693-0.954) and external (AUC = 0.773, 95% CI 0.647-0.902) test cohorts, and the DCA also suggested the radiomics-clinical model had the highest net benefit.
CONCLUSIONS: The radiomics-clinical model could be applied to the NPVR prediction of patients with uterine fibroids treated by HIFU to provide an objective and effective method for selecting potential patients who would benefit from the treatment mostly.
摘要：
背景：非灌注体积比（NPVR）的预测对于选择可能从超声引导的高强度聚焦超声（HIFU）治疗中受益的子宫肌瘤患者至关重要，因为它降低了治疗失败的风险。本研究的目的是通过机器学习，基于T2加权磁共振成像（T2MRI）影像组学特征结合临床参数，构建预测NPVR的最佳模型。
方法：这项回顾性研究是对来自两个中心的223例诊断为子宫肌瘤的患者进行的。来自一个中心的患者被分配到一个训练队列（n=122）和一个内部测试队列（n=46），来自其他中心的数据(n=55)用作外部测试队列.在训练队列中采用最小绝对收缩和选择算子（LASSO）算法进行特征选择。采用支持向量机(SVM)构建影像组学模型,临床模型,以及用于NPVR预测的影像组学临床模型，分别。曲线下面积(AUC)和决策曲线分析(DCA)评价模子的猜测效度和临床有用性，分别。
结果：从T2MRI中提取了851个放射学特征，其中7个影像组学特征被筛选为NPVR预测相关的影像组学特征。结合影像组学特征和临床参数的影像组学临床模型在内部（AUC=0.824，95％CI0.693-0.954）和外部（AUC=0.773，95％CI0.647-0.902）测试队列中均显示出最佳的预测性能。DCA还提示影像组学-临床模型的净获益最高.
结论：影像组学-临床模型可应用于HIFU治疗子宫肌瘤患者的NPVR预测，为筛选最有可能从治疗中获益的潜在患者提供客观有效的方法。
方法：这项回顾性研究是对来自两个中心的223例诊断为子宫肌瘤的患者进行的。来自一个中心的患者被分配到一个训练队列（n=122）和一个内部测试队列（n=46），来自其他中心的数据(n=55)用作外部测试队列.在训练队列中采用最小绝对收缩和选择算子（LASSO）算法进行特征选择。采用支持向量机(SVM)构建影像组学模型,临床模型,以及用于NPVR预测的影像组学临床模型，分别。曲线下面积(AUC)和决策曲线分析(DCA)评价模子的猜测效度和临床有用性，分别。
结果：从T2MRI中提取了851个放射学特征，其中7个影像组学特征被筛选为NPVR预测相关的影像组学特征。结合影像组学特征和临床参数的影像组学临床模型在内部（AUC=0.824，95％CI0.693-0.954）和外部（AUC=0.773，95％CI0.647-0.902）测试队列中均显示出最佳的预测性能。DCA还提示影像组学-临床模型的净获益最高.
结论：影像组学-临床模型可应用于HIFU治疗子宫肌瘤患者的NPVR预测，为筛选最有可能从治疗中获益的潜在患者提供客观有效的方法。
