Machine learning models based on CT radiomics for predicting the outcome of neoadjuvant chemotherapy in locally advanced gastric cancer
10.3760/cma.j.cn115355-20240809-00386
- VernacularTitle:基于CT影像组学的机器学习模型预测局部进展期胃癌新辅助化疗效果
- Author:
Feng HAN
1
;
Yanyan WANG
;
Yan DU
;
Jiaming CHENG
;
Erjuan WANG
;
Ruirui SONG
Author Information
1. 山西省肿瘤医院 中国医学科学院肿瘤医院山西医院 山西医科大学附属肿瘤医院信息科，太原 030013
- Publication Type:Journal Article
- Keywords:
Stomach neoplasms;
Neoadjuvant chemotherapy;
Machine learning;
Tomography, X-ray computed
- From:
Cancer Research and Clinic
2025;37(1):1-7
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the value of machine learning models based on CT radiomics for predicting the outcome of neoadjuvant chemotherapy (NAC) in patients with locally advanced gastric cancer (LAGC).Methods:A retrospective case series study was conducted. A total of 279 LAGC patients receiving NAC before surgery in Shanxi Province Cancer Hospital from January 2018 to November 2020 were included. According to a ratio of 7∶3, all patients were randomly divided into the training set (196 cases) and the validation set (83 cases). According to the tumor regression grade (TRG), the pathological grade was divided into the good response of NAC (GR) group (TRG 0-1, 55 cases) and the poor response of NAC (PR) group (TRG 2-3, 224 cases). The clinicopathological data of patients were collected, such as age, gender, differentiation degree, clinical T and N staging, carcinoembryonic antigen (CEA), and carbohydrate antigen 199 (CA199) level. Radiomics features were extracted from the enhanced CT images in the vein phase, and the features were screened by 3-step dimensionality reduction. And then 5 machine learning algorithms including logistic regression (LR), naive bayes (NB), random forest (RF), support vector machine (SVM) and extreme gradient boosting (XGB) were applied to build prediction models based on the CT radiomics. The receiver operating characteristic (ROC) curve and the decision analysis (DCA) curve were drawn to evaluate the predictive performance and clinical benefit of each model on the outcome of NAC in patients with LAGC.Results:Among 196 patients in the training set, there were 39 cases in GR group and 157 cases in PR group; among 83 patients in the validation set, there were 16 cases in GR group and 67 cases in PR group. There were no statistically significant differences in clinicopathological data of patients between the training and validation sets, or between GR and PR groups in the training and validation sets (all P > 0.05). A total of 102 radiomics features were extracted from region of interest of CT images in the vein phase, and 6 key features were finally selected including original_firstorder_10Percentile, original_firstorder_RoubustMeanAbsoluteDeviation, original_glcm_Idmn, original_glcm_MCC, original_ngtdm_Busyness, original_ngtdm_Contrast; and there were statistically significant differences in 6 features between the GR and PR groups (all P < 0.05). LR, NB, RF, SVM and XGB machine learning algorithms were used to construct 5 prediction models based on the CT radiomics. The area under ROC curve for NAC prediction in the training set was 0.553, 0.709, 0.668, 0.772 and 0.790, respectively; in the validation set was 0.662, 0.622, 0.683, 0.752 and 0.784, respectively. The model constructed by XGB showed the best comprehensive performance, and its accuracy, sensitivity and specificity was 0.771, 0.562 and 0.821, respectively. In the DCA of 5 machine learning models in the training set, XGB-based model provided a higher net benefit. Conclusions:Machine learning models based on enhanced CT radiomics in the vein phase have a high predictive efficacy in the outcome of NAC in LAGC patients before surgery and it helps make clinical personalized treatment decisions.
