Scholarlink

Objective <p>To develop and validate a machine learning model integrating preoperative contrast-enhanced CT radiomic features and clinical blood indicators for the auxiliary diagnosis of gastrointestinal stromal tumors (GIST).</p> Methods <p>This retrospective study collected data from 115 GIST patients and 202 non-GIST patients (including gastric cancer, gastric polyps, etc.) at Gansu Provincial Hospital between January 2017 and December 2024. All patients were randomly divided into training, validation, and test sets in a 7:2:1 ratio. Clinical features were identified using univariate and multivariate logistic regression analysis from a panel of 42 blood indicators; key radiomic features were extracted using LASSO regression from 851 features derived from contrast-enhanced CT images. Six machine learning algorithms were employed to construct clinical models, radiomics models, and a combined model, respectively. Model performance was evaluated using metrics including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Model interpretability and clinical utility were assessed using SHapley Additive exPlanations (SHAP) values, forest plots, and decision curve analysis (DCA).</p> Results <p>Multivariate analysis identified five clinical predictive indicators: DBIL, ALP, CEA, CA199, and PG I/PG II. LASSO regression selected four radiomic features: Sphericity, SurfaceVolumeRatio, etc. In the test set, both the radiomics model and the combined model demonstrated excellent diagnostic performance. The LightGBM radiomics model and the LightGBM combined model both achieved a high AUC (0.964 &amp;0.959). The clinical model showed relatively limited diagnostic value ( AUC: 0.745). SHAP analysis revealed that the radiomic feature Sphericity contributed the most to the combined model's decisions. The forest plot confirmed all selected features as significant predictors, although the feature importance rankings differed between interpretation methods. The decision curve analysis (DCA) demonstrated that both the LightGBM radiomics model and the LightGBM combined model offer significant clinical net benefit across a variety of decision thresholds, as evidenced by the algorithm's efficiency and accuracy in medical data analysis.</p> Conclusion <p>This study successfully developed machine learning models based on contrast-enhanced CT radiomics and blood indicators. The radiomics model and the combined model exhibited high diagnostic accuracy in distinguishing between benign and malignant lesions, as demonstrated in studies involving breast cancer, prostate lesions, and lung nodules. value for GIST, showing promise as effective preoperative auxiliary diagnostic tools to support clinical decision-making. Furthermore, the integration of multiple interpretation methods provided in-depth insights into the models' decision-making mechanisms, offering a solid foundation for the clinical translation of this technology.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning model integrating CT radiomics and clinical blood indicators for auxiliary diagnosis of gastrointestinal stromal tumors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月22日

查看全文
  • Jing Yang,
  • Aijia Yang,
  • Pei Liu,
  • Xiaowei Zhu,
  • Yi Liu,
  • Zhiping Ma,
  • Zhengjie Xu,
  • Long Zhao,
  • Lili Wang,
  • Hui Cai

摘要

Objective

To develop and validate a machine learning model integrating preoperative contrast-enhanced CT radiomic features and clinical blood indicators for the auxiliary diagnosis of gastrointestinal stromal tumors (GIST).

Methods

This retrospective study collected data from 115 GIST patients and 202 non-GIST patients (including gastric cancer, gastric polyps, etc.) at Gansu Provincial Hospital between January 2017 and December 2024. All patients were randomly divided into training, validation, and test sets in a 7:2:1 ratio. Clinical features were identified using univariate and multivariate logistic regression analysis from a panel of 42 blood indicators; key radiomic features were extracted using LASSO regression from 851 features derived from contrast-enhanced CT images. Six machine learning algorithms were employed to construct clinical models, radiomics models, and a combined model, respectively. Model performance was evaluated using metrics including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Model interpretability and clinical utility were assessed using SHapley Additive exPlanations (SHAP) values, forest plots, and decision curve analysis (DCA).

Results

Multivariate analysis identified five clinical predictive indicators: DBIL, ALP, CEA, CA199, and PG I/PG II. LASSO regression selected four radiomic features: Sphericity, SurfaceVolumeRatio, etc. In the test set, both the radiomics model and the combined model demonstrated excellent diagnostic performance. The LightGBM radiomics model and the LightGBM combined model both achieved a high AUC (0.964 &0.959). The clinical model showed relatively limited diagnostic value ( AUC: 0.745). SHAP analysis revealed that the radiomic feature Sphericity contributed the most to the combined model's decisions. The forest plot confirmed all selected features as significant predictors, although the feature importance rankings differed between interpretation methods. The decision curve analysis (DCA) demonstrated that both the LightGBM radiomics model and the LightGBM combined model offer significant clinical net benefit across a variety of decision thresholds, as evidenced by the algorithm's efficiency and accuracy in medical data analysis.

Conclusion

This study successfully developed machine learning models based on contrast-enhanced CT radiomics and blood indicators. The radiomics model and the combined model exhibited high diagnostic accuracy in distinguishing between benign and malignant lesions, as demonstrated in studies involving breast cancer, prostate lesions, and lung nodules. value for GIST, showing promise as effective preoperative auxiliary diagnostic tools to support clinical decision-making. Furthermore, the integration of multiple interpretation methods provided in-depth insights into the models' decision-making mechanisms, offering a solid foundation for the clinical translation of this technology.