Application of a spectral CT-based radiomics machine learning model and nomogram for preoperative identification of KRAS gene status in colorectal cancer

LI Zemao; MA Ruhang; WANG Yajing; CHEN Weibin

doi:10.19300/j.2025.L21648

Abstract：Objective To explore the diagnostic performance of a spectral CT-based radiomics machine learning model and nomogram for preoperatively identifying the KRAS gene status in patients with colorectal cancer(CRC).Methods A total of 137 CRC patients who underwent KRAS mutation detection and preoperative spectral CT examination were retrospectively included(70 cases with KRAS wild type and 67 cases with KRAS mutant type).They were randomly divided into a training set(95 cases)and a test set(42 cases)in a 7∶3 ratio.Tumor region of interest(ROI)was delineated on venous-phase 70 keV monochromatic enhanced CT images,and radiomics features were extracted and selected.A radiomics score(Rad-score)was calculated using least absolute shrinkage and selection operator(LASSO)regression.Six models were established including three radiomics models based on support vector machine(SVM),extreme gradient boosting(XGBoost),and logistic regression(LR),as well as three combined models integrating spectral CT imaging features with the Rad-score.Model performance was evaluated using the area under the receiver operating characteristic(ROC)curve(AUC),and compared using the Delong test.A radiomics nomogram was constructed based on the Rad-score and validated in the test set.Calibration curves,decision curve analysis(DCA),and clinical impact curves were used to assess calibration,clinical net benefit,and clinical utility.Results A total of 8 radiomics features and 1 spectral parameter were selected.In the test set,the LR-based combined model demonstrated the best performance,with an AUC of 0.891,outperforming the combined models based on SVM(AUC=0.796),XGBoost(AUC=0.787),and LR(AUC=0.812)(all P＜0.05),as well as the combined models based on SVM(AUC=0.889)and XGBoost(AUC=0.873)(both P＜0.05).The nomogram model achieved AUCs of 0.987 and 0.916 in the training and test sets,respectively.The calibration curve showed good agreement in the training set,while performance in the test set was slightly lower.DCA and clinical impact curves demonstrated that the nomogram provided favorable clinical net benefit and utility.Conclusion The LR-based model and nomogram,constructed using venous-phase spectral CT and radiomics features,offer valuable preoperative insights into KRAS gene status in CRC patients and may serve as a reference for clinical decision-making.