Ki-67 expression, a critical biomarker for tumor aggressiveness and proliferation in invasive breast cancer, is traditionally assessed via invasive biopsies, which suffer from sampling variability and limit serial monitoring. This study aimed to develop radiomics models using ultrasound (US) and mammography (MG) features to predict Ki-67 expression, hypothesizing that dual-modality integration improves accuracy over single-modality approaches. A retrospective study was performed on consecutive patients diagnosed with invasive breast cancer at Second Affiliated Hospital of Fujian University of Traditional Chinese Medicine from January 2017 to May 2024. A total of 206 eligible patients were ultimately enrolled, including 98 in the low Ki-67 group (Ki-67 ≤ 20%) and 108 in the high Ki-67 group (Ki-67 > 20%). Radiomic features were extracted from US and MG images using Pyradiomics and refined via least absolute shrinkage and selection operator (LASSO) regression. The dataset was split into training and test sets at a 7: 3 ratio using stratified sampling to ensure proportional balance across classes. Radiomic features were extracted from US and MG images using Pyradiomics and refined via least absolute shrinkage and selection operator (LASSO) regression. For tumor segmentation, two investigators manually delineated regions of interest (ROIs) along tumor boundaries on US and MG images using ITK-SNAP software. To ensure robustness, intraclass correlation coefficient (ICC) analysis was performed on features extracted from annotations by two independent observers for 50 random cases (including both US and MG studies). Features with ICC ≥ 0. 75 were retained. Four machine learning classifiers (support vector machine SVM, logistic regression LR, Random Forest, XGBoost) were constructed, with clinical metrics integrated into model development. Model validation was conducted via three-round 5-fold cross-validation. The key features were identified using LASSO and interpreted via Shapley Additive exPlanations (SHAP) analysis. Clinical utility was assessed using decision curve analysis (DCA). High Ki-67 expression (> 20%) was significantly associated with higher histologic grade (Grade III: 76. 9% vs. 23. 5%, p < 0. 001), larger tumor size (2. 1–5. 0 cm: 61. 1% vs. 41. 8%, p = 0. 019), ER/PR negativity (p < 0. 001), HER2 positivity (p = 0. 010), and aggressive molecular subtypes (Luminal B, HER2+, Triple-negative; p < 0. 001). After feature selection, 12 US, 17 MG, and 42 combined features were retained. The combined model significantly outperformed single-modality models in test-set 3-round 5-fold cross-validation, achieving the highest area under the curve (0. 882 vs. 0. 748 for US and 0. 771 for MG, p < 0. 05), with balanced sensitivity (83. 0 ± 9. 3%) and specificity (73. 0 ± 7. 6%). Texture-based features (e. g. , usSquareglszmSizeZoneNonUniformityNormalized) were critical predictors by SHAP analysis. DCA demonstrated the combined model offered the greatest net benefit across threshold probabilities (20–60%), confirming its superior clinical utility. It suggests that the combined model has potential clinical utility for guiding personalized strategies, such as adjusting therapy intensity based on predicted Ki-67 levels, which requires further prospective validation. Integrating US and MG radiomics significantly improves the non-invasive prediction of Ki-67 expression in invasive breast cancer compared to single-modality models. However, the study is limited by its single-center design, and multi-center validation with external cohorts is needed to enhance generalizability. Future work should also explore hybrid models with optimized integration of clinical features and prospective evaluation to confirm clinical utility.
Xu et al. (Thu,) studied this question.