中国实用外科杂志

中国实用外科杂志 ›› 2022, Vol. 42 ›› Issue (12): 1401-1407.DOI: 10.19538/j.cjps.issn1005-2208.2022.12.12

• 论著 • 上一篇    下一篇

MSCT影像组学结合机器学习预测直径2~5 cm胃胃肠间质瘤危险度分级研究

吴树剑a，俞咏梅a，范莉芳b，徐静雅a，任    超a，徐争元b，徐晓燕b   

  1. 皖南医学院 a.第一附属医院（弋矶山医院）医学影像中心  b.医学影像学院，安徽芜湖 241000
  • 出版日期:2022-12-01 发布日期:2022-12-27

  • Online:2022-12-01 Published:2022-12-27

摘要： 目的    探讨多层螺旋CT（MSCT）影像组学结合不同机器学习（ML）算法预测直径2~5 cm胃胃肠间质瘤（GIST）危险度分级的价值。方法    回顾性分析115例（共115枚）直径2~5 cm术后病理学检查证实为胃GIST病人的完整资料，按照美国国家卫生研究院（NIH）分级标准，分为低风险组67例（极低危险度5例，低危险度62例）和高风险组48例（均为中危险度）。按7∶3比例将病人随机分为训练集（n=80）和验证集（n=35）。利用ITK-SNAP软件分割图像，利用AK软件提取影像组学特征，并降维构建影像组学标签评分（rad-score）。采用单因素及多因素Logistic回归筛选独立危险因素，利用支持向量机（SVM）、最近邻算法（KNN）、极端梯度增强机（XGBoost）3种机器学习算法构建预测模型，并筛选最优模型为输出模型。采用受试者操作特征（ROC）曲线下面积（AUC）评价模型的效能，使用临床决策曲线（DCA）评价输出模型的临床适用性。结果    单因素分析消化道出血、血管样强化、分叶征、坏死、长径、短径、强化特征、静脉期不均匀率（SHRTv）及rad-score差异有统计学意义（P＜0.05）。多因素分析血管样强化、强化特征及rad-score为独立危险因素。3种ML算法构建预测模型的AUC分别为训练集0.912、0.865、0.887，验证集0.897、0.845、0.865，SVM为最优模型。DCA显示SVM算法构建预测模型在训练集及验证集均有较高的临床适用性。结论    MSCT影像组学结合机器学习能够无创、准确地预测胃GIST（2~5 cm）危险度分级，为临床外科医师制定治疗方案提供参考依据。

关键词: 多层螺旋CT, 影像组学, 机器学习, 胃, 胃肠间质瘤, 危险度分级

Abstract: MSCT radiomics combined with machine learning to predict the risk grading of gastric stromal tumors with a diameter of 2-5 cm        WU Shu-jian*, YU Yong-mei, FAN Li-fang, et al. *Medical Imaging Center of Yijishan Hospital Affiliated to Wannan Medical College, Wuhu 241000, China
Corresponding author: YU Yong-mei, E-mail:yjsyym131@163.com
Abstract    Objective    To investigate the value of multi-slice spiral CT (MSCT） radiomics combined with different machine learning（ML）algorithms in predicting the risk classification of gastric stromal tumors (GIST） with a diameter of 2-5 cm. Methods    The complete data of 115 patients (a total of 115） with GIST confirmed by pathology after 2-5 cm operation were retrospectively analyzed, according to the National Institutes of Health (NIH） classification criteria, 67 cases were divided into a low-risk group (5 cases were very low-risk, 62 cases were low risk） and 48 cases were high risk group(both medium risk）. The patients were randomly divided into training set (n=80） and validation set (n=35） according to 7∶3 ratio. Images were segmented by ITK-SNP software, radiomics features were extracted by AK software, and radiomics tag score (rad-score） was constructed by dimensionality reduction. Univariate and multivariate Logistic regression was used to screen the independent risk factors. Three machine learning algorithms, support vector machine (SVM）, k-nearest neighbor (KNN）, and extreme gradient boosting (XGBoost）, were used to construct the prediction model, and the optimal model was selected as the output model. The area under the receiver operating characteristic (ROC） curve (AUC） was used to evaluate the efficiency of the model, and the clinical decision curve (DCA） was used to assess the clinical applicability of the output model. Results    Univariate analysis showed significant differences in gastrointestinal bleeding, vascular enhancement, lobulation sign, necrosis, long diameter, short diameter, enhancement characteristics, standardized heterogeneous rate of tumor of venous (SHRTv） and rad-score (P＜0.05）. Multivariate analysis showed that vascular enhancement，enhancement characteristics and rad-score were independent risk factors. The AUC of the prediction model constructed by the three ML algorithms were 0.912, 0.865, 0.887 for the training set, 0.897, 0.845, 0.865 for the validation set, and the SVM was the optimal model. DCA shows that the prediction model constructed by SVM algorithm has high clinical applicability in both training and validation sets. Conclusion    MSCT radiomics combined with machine learning can noninvasively and accurately predict the risk grade of GIST（2-5 cm）, which provides a reference for clinical surgeons to make treatment plans.

Key words: multi-slice spiral CT, radiomics, machine learning, stomach, gastrointestinal stromal tumor, the risk classification