A hybrid radiomics framework integrating genetic algorithm-optimized random forest for preoperative identification of Luminal B breast cancer and Ki-67 prediction: A multicenter study

BACKGROUND: Preoperative identification of Luminal B breast cancer remains a clinical challenge. This study aimed to develop an ultrasound radiomics framework integrating tumoral and peritumoral information for preoperative identification of Luminal B subtype and prediction of Ki-67 status. METHODS: We retrospectively analyzed 1,944 patients from three centers. The development cohort from Centers One and Two was divided by stratified sampling into a training set (n = 1,434) and an internal test set (n = 253), and an independent cohort from Center Three (n = 257) was used for external validation. Lesion-containing ROIs were processed using deep learning-assisted segmentation and standardized for downstream analysis. Radiomic features were extracted, and a genetic algorithm (GA) was coupled with a random forest (RF) classifier to construct two models: one for Luminal B classification and another for predicting Ki-67 expression. RESULTS: The combined tumor-peritumoral model achieved the highest performance, with the Luminal B classifier showing AUCs of 0.876 (training), 0.693 (test), and 0.786 (external validation). The Ki-67 prediction model yielded AUCs of 0.890 (training) and 0.858 (test), though external validation (AUC=0.661) was limited by dataset distribution. The Delong test confirmed that combined ROIs significantly outperformed tumor-only models, with NRI and IDI tests further validating the added value of peritumoral features. CONCLUSIONS: Ultrasound radiomics integrating tumoral and peritumoral regions can support the preoperative identification of Luminal B breast cancer, and peritumoral region analysis significantly enhances predictive performance. The framework also shows potential for predicting Ki-67 status within this subtype.