Related Experiment Video
Updated: Aug 30, 2026
Gene Regulation and Targeted Therapy in Gastric Cancer Peritoneal Metastasis: Radiological Findings from Dual Energy CT and PET/CT
Published on: January 22, 2018
Prediction of human epidermal growth factor receptor 2 expression in gastric carcinoma using whole-tumor
Xue-Qin Wang1, Hao Wen1, Qiu-Hong Lu1
1Key Laboratory of Intelligent Processing and Applications of Medical Imaging Big Data, Chongqing Municipal Health Commission, and Department of Radiology, The Second Affiliated Hospital of Chongqing Medical University, Chongqing, China.
Purpose:
To investigate feasibility of combination of whole-tumor subregion-based CT habitat and conventional radiomics to improve preoperative identifying human epidermal growth factor receptor 2 (HER2) expression in gastric carcinoma (GC).
Methods:
328 patients with GC undergoing preoperative contrast-enhanced CT were retrospectively included from two centers. Patients from Center 1 (n = 278) were randomly divided into training (n = 222) and internal-validation (n = 56) cohorts, and an external-validation cohort comprised 50 patients from Center 2. Conventional CT radiomics features were extracted from whole-tumor. Habitat radiomics features were from three habitat subregions of whole-tumor using K-means clustering. The habitats-based and conventional radiomics models, and the combined model by integrating previous features for predicting HER2 status were created using Support Vector Machine (SVM) classifiers followed with eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). Models' performance was evaluated using area under the receiver operating characteristic curve (AUC) and Delong test.
Results:
For SVM classifiers, the combined model showed better predictability than the habitats-based or conventional model (AUC: 0.956 vs. 0.911 or 0.792, 0.847 vs. 0.835 or 0.771, and 0.725 vs. 0.685 or 0.614) in training, internal- and external-validation cohorts, which was confirmed using XGBoost (AUC: 0.784 vs. 0.754 or 0.702, and 0.745 vs. 0.696 or 0.638) and LightGBM (AUC: 0.804 vs. 0.748 or 0.685, and 0.708 vs. 0.699 or 0.553) classifiers in internal- and external-validation cohorts, respectively (p < 0.05 for most comparisons).
Conclusion:
The combined model could serve as a predictive biomarker of HER2 expression prediction in GC.
