本研究中，回顾性方法被采用，且GISTs危险度分级以及KI67指数高低

本研究中，回顾性方法被采用，且GISTs危险度分级以及KI67指数高低的预测模型被构建基于多中心CT（computed tomography）影像数据并利用多分类器影像组学技术，有潜力帮助临床医生进行患者分层，精准治疗。<br>研究共分为以下两个部分。<br><br><br>第一部分 <br>基于CT影像组学方法预测胃肠道间质瘤危险度分级<br><br>目的：<br>肿瘤复发风险可借助胃肠道间质瘤(GISTs)的病理危险度分级判断。<br>CT（computed tomography）是目前GISTs诊疗的主要手段。<br><br>本研究被设计以试图开发并验证基于CT影像组学技术的胃肠道间质瘤(GISTs)危险度分级术前预测模型。<br><br>方法：<br>回顾性分析被针对分别来自3家医院（训练集180例，外部测试集144例）的经手术病理证实的GIST患者324例开展。<br><br>应用基于MATLAB的IBEX软件包在动脉期、静脉期CT图像上沿着肿瘤轮廓内缘逐层勾画感兴趣区（regions of interest, ROI），提取GISTs的影像组学特征，逐步特征降维。<br><br><br>训练集使用3个分类算法（Logistics回归、SVM与随机森林）分别建立GIST危险度分级预测模型，利用混淆矩阵相关参数及受试者工作特征曲线（receiver operating characteristiccurve, ROC）评估与比较模型效能，并利用外部数据进行验证。<br><br>结果：<br>In the binary classification prediction of GISTs risk, Logistic regression has AUC of 0.84 and 0.85 in the training set and external validation set, respectively. The AUC of SVM in the training set and external validation set are 0.81 and 0.80, respectively. Random forest has AUC in the training set and external validation set. They are 0.88 and 0.90 respectively. <br><br><br>The random forest model has the best performance in both the training set and the test set, and its generalization ability is good. <br><br>In the three classifications of GISTs risk, the logistic regression training set has a macro-average AUC of 0.85 and a micro-average AUC of 0.85. In the external verification, the macro-average AUC is 0.75 and the micro-average AUC is 0.77. <br><br>SVM in the training set has a macro average AUC of 0.50 and a micro average AUC of 0.61. <br>The macro average AUC in the external verification set was 0.50, and the micro average AUC was 0.60. <br><br>The random forest in the training set has a macro-average AUC of 0.79 and a micro-average AUC of 0.81. In external verification, the macro-average AUC is 0.83 and the micro-average AUC is 0.83. SVM performs the worst, and random forest is the most stable. <br><br>Conclusion: <br>Different machine learning models based on CT radiomics can effectively predict GIST risk classification. Among them, the radiomics model based on the random forest algorithm has the highest predictive power and strong generalization ability, and it has been confirmed in external test data. It has the potential to help clinicians predict the GIST risk level before surgery, and can help clinicians stratify patients and treat them accurately. <br> <br>[Keywords] gastrointestinal stromal tumors; risk classification; radiomics;

In this study, retrospective methods were used, and GISTs risk classification and ki67 index prediction models were constructed based on multi-center CT (computed tomography) image data and utilized multi-classifier imaging histology techniques, with the potential to help clinicians stratify patients and treat them with precision.<br>The study is divided into the following two parts.<br><br>The first part. <br>Based on CT imaging histology, the risk rating of gastrointestinal mesothelioma was predicted.<br><br>Objective:<br>The risk of tumor recurrence can be determined by the classification of pathological risk of gastrointestinal mesothelioma (GISTs).<br>CT (computed tomography) is currently the main means of GISTs diagnosis and treatment.<br><br>This study was designed to attempt to develop and validate a preoperative prediction model for the classification of the risk of gastrointestinal mesothelioma (GISTs) based on CT imaging histology techniques.<br><br>Method:<br>The retrospective analysis was conducted on 324 surgically pathologically confirmed GIST patients from 3 hospitals (180 training sets and 144 external test sets).<br><br>MatLAB-based IBEX software packages are used to map the area of interest (regions of interest, ROI) layer by layer along the inner edge of tumor contours on arterial and intravenous CT images, extracting the image histological characteristics of GISTs and gradually reducing the dimensional characteristics.<br><br>The training set uses three classification algorithms (Logistics regression, SVM and random forest) to establish a GIST risk grading prediction model, evaluate and compare model effectiveness using the confusion matrix-related parameters and the subject's working characteristic curve (receiver operating characteristics curve, ROC), and verify it with external data.<br><br>Results:<br>In GISTs risk ii classification prediction, Logistic regression AAUC was 0.84 and 0.85 in training set and external verification set, 0.81 and 0.80 in training set and external verification set, and random forest AAUC in training set and external verification set, respectively.<br><br>The random forest model is the best performance in both the training set and the test set, and the generalization ability is good.<br><br>In the GISTs risk three categories, the logistic regression training concentration macro average AAUC is 0.85, the micro-average AUC is 0.85, the external verification macro average AUC is 0.75, and the micro-average AAUC is 0.77.<br><br>SVM has an average AUC of 0.50 and a micro-average AUC of 0.61 in training concentration.<br>The average AUC of the external verification centralized macro is 0.50 and the micro-average AUC is 0.60.<br><br>Random Forest has an average AUC of 0.79, a micro-average of 0.81, an AUC of 0.83 and a micro-average of 0.83 in external verification. SVM performed worst and random forests were the most stable.<br><br>Conclusion:<br>Different machine learning models based on CT radiology can effectively predict GIST risk classification, in which the radiological model based on random forest algorithm has the highest prediction effectiveness, generalization ability, and it is proved in external test data, has the potential to help clinicians predict GIST risk level before surgery, can help clinicians to stratify patients, precision treatment.<br> <br>Keywords: gastrointestinal mesothelioma; risk classification; radiology;

In this study, retrospective method was adopted, and the prediction model of GISTs risk classification and Ki67 index was constructed. Based on multi center CT (computed tomography) image data and using multi classifier imaging technology, it has the potential to help clinicians to stratify patients and treat patients accurately.<br>The research is divided into the following two parts.<br>Part one<br>Prediction of risk classification of gastrointestinal stromal tumors based on CT imaging<br>Objective: To study the clinical effect of the method<br>The risk of tumor recurrence can be judged by the pathological risk classification of gastrointestinal stromal tumors (GISTs).<br>Computed tomography (CT) is the main method of diagnosis and treatment of GISTs.<br>This study was designed to attempt to develop and validate a preoperative prediction model for risk classification of gastrointestinal stromal tumors (GISTs) based on CT imaging.<br>method:<br>A retrospective analysis was conducted on 324 patients with GIST confirmed by surgery and pathology from three hospitals (180 cases in the training set and 144 cases in the external test set).<br>The ibex software package based on MATLAB was used to delineate the regions of interest (ROI) along the inner edge of tumor contour in arterial phase and venous phase CT images, and the imaging characteristics of GISTs were extracted, and the feature dimension was gradually reduced.<br>In the training set, three classification algorithms (logistic regression, SVM and random forest) were used to establish the gist risk classification prediction model. The relevant parameters of confusion matrix and receiver operating characteristic curve (ROC) were used to evaluate and compare the effectiveness of the model, and external data were used to verify the model.<br>result:<br>The AUC of logistic regression in training set and external verification set were 0.84 and 0.85, respectively. AUC of SVM in training set and external verification set were 0.81 and 0.80, respectively. AUC of random forest in training set and external verification set were 0.88 and 0.90, respectively.<br>The performance of random forest model is the best in both training set and test set, and has good generalization ability.<br>In the three categories of GISTs risk, the macro average AUC and micro average AUC of logistic regression training set were 0.85 and 0.85 respectively. In the external verification, the macro average AUC was 0.75 and the micro average AUC was 0.77.<br>In the training set, the macro average AUC of SVM is 0.50, and the micro average AUC is 0.61.<br>In the external validation set, the macro average AUC was 0.50, and the micro average AUC was 0.60.<br>In the training set, the macro average AUC and micro average AUC of random forest were 0.79 and 0.81 respectively. In the external verification, the macro average AUC was 0.83 and the micro average AUC was 0.83. SVM is the worst and random forest is the most stable.<br>Conclusion<br>Different machine learning models based on CT radioomics can effectively predict the gist risk classification. Among them, the radioomics model based on random forest algorithm has the highest prediction efficiency and strong generalization ability. Moreover, it has been confirmed in the external test data. It has the potential to help clinicians predict the gist risk level before operation, and can help clinicians to stratify patients and make accurate treatment.<br>[Key words] gastrointestinal stromal tumor; risk classification; radioomics;<br>