| 张源,侯娟,朱玉才,阿不都热苏力·吐尔孙,郭辉.基于机器学习门静脉期CT影像组学模型预测肝泡型棘球蚴病术后并发症[J].中国医学影像技术,2025,41(9):1535~1539 |
| 基于机器学习门静脉期CT影像组学模型预测肝泡型棘球蚴病术后并发症 |
| Portal venous phase CT radiomics model based on machine learning for predicting postoperative complications of hepatic alveolar echinococcosis |
| 投稿时间：2025-04-21 修订日期：2025-08-18 |
| DOI：10.13929/j.issn.1003-3289.2025.09.017 |
| 中文关键词: 棘球蚴病,肝|手术后并发症|体层摄影术,X线计算机|机器学习|影像组学 |
| 英文关键词:echinococcosis, hepatic|postoperative complications|tomography, X-ray computed|machine learning|radiomics |
| 基金项目:新疆人工智能影像辅助诊断重点实验室开放课题(XJRGZN2024016)。 |
|
| 摘要点击次数: 89 |
| 全文下载次数: 32 |
| 中文摘要: |
| 目的 观察基于机器学习(ML)门静脉期CT影像组学模型预测肝泡型棘球蚴病(HAE)术后并发症的价值。方法 回顾性纳入265例HAE,按7∶3比例将其随机分为训练集(n=185,含106例发生术后并发症)与验证集(n=80,含40例发生术后并发症)。基于门静脉期CT图像分割HAE病灶,提取、筛选影像组学特征;分别以5种ML算法构建模型,比较其预测术后并发症的效能。结果 5种ML影像组学模型中,以支持向量机(SVM)模型预测训练集及验证集中HAE术后并发症的整体表现最优;DeLong检验显示,训练集中,SVM模型的曲线下面积(AUC)均显著大于逻辑回归(LR)、K邻近法(KNN)及多层感知器(MLP)模型(P均＜0.001),在验证集中则显著大于自适应增强(AdaBoost)模型(P=0.007)。决策曲线分析显示,SVM模型临床净收益最高。结论 基于ML算法、尤其SVM算法的门静脉期CT影像组学模型可有效预测HAE术后并发症。 |
| 英文摘要: |
| Objective To observe the value of portal venous phase CT radiomics model based on machine learning (ML) for predicting postoperative complications of hepatic alveolar echinococcosis (HAE). Methods Totally 265 HAE patients were retrospectively enrolled and randomly divided into training set (n=185, including 106 cases with postoperative complications) and validation set (n=80, including 40 cases with postoperative complications) at a ratio of 7∶3. Based on portal venous phase CT images, HAE lesions were segmented, and radiomics features were extracted and screened. Totally 5 ML algorithms were used to construct models, and their performance for predicting postoperative complications were compared. Results Among 5 ML radiomics models, support vector machine (SVM) model had the best overall performance for predicting postoperative complications of HAE in both training and validation sets. DeLong test showed that in training set, the area under the curve (AUC) of SVM model was significantly higher than that of logistic regression (LR), K-nearest neighbor (KNN) and multilayer perceptron (MLP) models (all P＜0.001), while in validation set, the AUC of SVM model was significantly higher than that of adaptive boosting (AdaBoost) model (P=0.007). Decision curve analysis indicated that SVM model had the highest clinical net benefit. Conclusion Portal venous phase CT radiomics model based on ML algorithms, especially SVM algorithm, could effectively predict postoperative complications of HAE. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
