Preoperative prediction of Ki-67 expression in breast cancer using a combined radiomics model based on 18F-FDG PET/CT

This study aimed to develop and validate a noninvasive predictive model integrating conventional metabolic parameters and radiomic features derived from 18F-fluorodeoxyglucose positron emission tomography (PET)/computed tomography (CT) for the preoperative assessment of proliferation marker Ki-67 expression in patients with breast cancer (BC). This retrospective study adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis, Extended for Artificial Intelligence guidelines and included 119 patients with BC who underwent preoperative 18F-fluorodeoxyglucose PET/CT. Patients were randomly assigned to a training cohort (n = 95) and an independent testing cohort (n = 24) at a 4:1 ratio. Radiomics scores (Radscores) were constructed separately for CT and PET using logistic regression models weighted by features selected via least absolute shrinkage and selection operator regression. Five machine learning algorithms - logistic regression, support vector machine (SVM), random forest, extreme gradient boosting, and k-nearest neighbors - were employed to build combined models integrating significant clinical, imaging, and serum tumor marker features with CT-Radscore and PET-Radscore. Model performance was evaluated in the testing set using the area under the receiver operating characteristic curve (AUC), precision-recall curves, and decision curve analysis. The DeLong test was used to compare AUCs between the optimal combined model and individual predictors. Model interpretability was enhanced using SHapley Additive exPlanations to quantify feature contributions at both global and case-specific levels. In the testing set, CT-Radscore and PET-Radscore achieved AUCs of 0.785 (95% confidence interval [CI]: 0.592, 0.977) and 0.806 (95% CI: 0.625, 0.986), respectively. The SVM-based combined model incorporating maximum standardized uptake value, total lesion glycolysis, CT-Radscore, and PET-Radscore demonstrated superior performance, yielding an AUC of 0.847 (95% CI: 0.684, 1.000), which outperformed all single-variable predictors. Precision-recall curves confirmed its robust discriminative ability, while decision curve analysis showed that the SVM model provided greater net clinical benefit across a wide range of threshold probabilities compared with any individual predictor. SHapley Additive exPlanations analysis identified PET-Radscore as the most influential feature, followed by CT-Radscore, maximum standardized uptake value, and total lesion glycolysis. A combined SVM model integrating PET/CT-derived metabolic parameters and radiomic features showed preliminary ability to noninvasively predict high proliferation marker Ki-67 expression in BC. This supports the potential utility of multimodal radiomics in preoperative assessment of tumor proliferative status, pending further validation.