Interpretable prediction of Ki-67 expression in breast cancer through integrated photoacoustic radiomics and clinical parameters

Purpose: This study aimed to develop a noninvasive method for preoperatively predicting Ki-67 expression in breast cancer (BC) by integrating radiomics features derived from photoacoustic/ultrasound (PA/US) imaging with clinical factors. Materials and methods: A total of 223 patients with pathologically confirmed BC underwent PA/US imaging before surgery. Radiomics features were extracted from tumor regions, standardized, and selected using statistical testing, correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. Ten machine learning algorithms were compared, and the best-performing classifier was used to construct the radiomics model. Clinical variables significantly associated with Ki-67 expression were then incorporated to build a combined model. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, calibration analysis, and decision curve analysis. Results: In the test set, the clinical, radiomics, and combined models achieved areas under the ROC curve (AUCs) of 0.722, 0.826, and 0.849, respectively. The combined model also demonstrated higher sensitivity (0.957) and negative predictive value (0.867), indicating strong capability in identifying patients with high Ki-67 expression. SHAP analysis confirmed that texture- and intensity-related imaging features played important roles in model prediction. Conclusion: PA/US-based radiomics provides a quantitative, interpretable, and clinically valuable approach for assessing tumor proliferation in BC. The combined model exhibited superior accuracy over single-modality models and shows promise as an effective supplementary tool for individualized preoperative evaluation of Ki-67 expression.