Non-contrast CT radiomics extreme gradient boosting(XGBoost)model for predicting acute necrotic collection around acute pancreatitis
10.13929/j.issn.1003-3289.2025.02.021
- VernacularTitle:平扫CT影像组学极限梯度提升(XGBoost)模型预测急性胰腺炎周围坏死物积聚
- Author:
Yuyu YU
1
;
Hanlin ZHU
;
Peiying WEI
;
Haifeng ZHANG
;
Bo FENG
Author Information
1. 杭州市第九人民医院放射科,浙江 杭州 311225;西湖大学医学院附属杭州市第一人民医院放射科,浙江 杭州 310006
- Publication Type:Journal Article
- Keywords:
pancreatitis;
necrotic collection;
radiomics
- From:
Chinese Journal of Medical Imaging Technology
2025;41(2):281-285
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of non-contrast CT radiomics extreme gradient boosting(XGBoost)model based on SHAP method for predicting acute necrotic collection(ANC)around acute pancreatitis(AP).Methods A total of 307 patients with initially clinically diagnosed AP were retrospectively enrolled.The optimal radiomics features of peripheral pancreatic tissue volume of interest(VOI)were extracted and screened based on automatic segmentation on the first non-contrast CT,and the evaluation results of modified CT severity index(MCTSI)score of AP severity based on first enhanced CT were recorded.The patients were divided into peripancreatic ANC group(ANC group)and acute peripancreatic fluid collection(APFC)group according to follow-up abdominal CT.XGBoost method was used to construct radiomics model,MCTSI model and combined model for predicting AP ANC based on the optimal radiomics features,MCTSI and their combination,respectively.The diagnostic efficacy of each model was evaluated using 5-fold cross-validation method,and the contribution of each variable to combined model was analyzed with SHAP method.Results Among 307 cases,there were 134 cases in ANC group and 173 in APFC group.Totally 6 optimal radiomics features were screened based on the first non-contrast CT.The area under the receiver operating characteristic curve(AUC)of radiomics model,MCTSI model and combined model was 0.936,0.693 and 0.917,respectively.The AUC of MCTSI model was lower than that of radiomics model and combined model(Z=-3.485,-2.824,both P＜0.01),while no significant difference of AUC was found between radiomics model and combined model(Z=-0.817,P=0.415).The contribution of optimal radiomics features to combined model were all higher than that of MCTSI score.Conclusion Non-contrast CT radiomics XGBoost model could effectively predict AP ANC.
