Enhancing prognostic accuracy in invasive breast cancer by combining contrast-enhanced ultrasound and clinical data: a multicenter retrospective study.

Background: Current predictive models for disease-free survival (DFS) in invasive breast cancer predominantly utilize clinical and pathological factors, with minimal incorporation of ultrasound (US) and contrast-enhanced ultrasound (CEUS) characteristics. This study aimed to establish a multimodal map integrating US, clinical features, and US data to enhance the prediction of DFS in invasive breast cancer. Methods: The study utilized three retrospective datasets obtained from three academic medical centers, covering the period from March 2014 to December 2022. Clinical data, gray scale US, and CEUS were assessed in 942 adult patients undergoing breast cancer resection. The training and internal test sets were supplied by The First Medical Center of the PLA General Hospital, while the external test sets were sourced from The Fourth Medical Center of the PLA General Hospital and the Specialist Medical Center of the Strategic Support Forces. The patients were followed up by phone or clinic visits. DFS was evaluated as a prognostic outcome. Cox regression analysis identified prognostic factors, leading to the construction of three nomograms. The model performance was evaluated using the C-index, time-dependent receiver operating characteristic (ROC) curve, calibration, decision curve analysis (DCA), integrated discrimination improvement (IDI), and net reclassification index (NRI). Results: A total of 942 patients were enrolled, with a mean age of 51.91 years [interquartile range (IQR), 44.25-58.69 years]. The patients were included with the median DFS of 36 months. Cox regression analysis identified menopausal status, body mass index (BMI), color Doppler flow imaging (CDFI), tumor size on CEUS, adjuvant/neoadjuvant chemotherapy, progesterone receptor (PR) status, and tumor-node-metastasis (TNM) staging as significant risk factors for invasive breast cancer. The nomogram combining US, CEUS, and clinical data demonstrated excellent predictive performance, achieving a C-index of 0.811 in the training set, 0.816 in the internal validation set, and 0.819 in the external validation set. Calibration curves confirmed that the predicted survival probabilities aligned closely with observed outcomes. Comparative analysis of ROC curves, IDI, NRI, and DCA confirmed that the integrated nomogram outperformed models based solely on US and clinical data or clinical data alone in predicting 24- and 36-month DFS. Conclusions: The integration of CEUS and clinical factors for non-invasive DFS prediction improves personalized risk stratification, minimizing unnecessary interventions for low-risk patients and ensuring adequate monitoring for high-risk individuals. Additional prospective validation is required to establish its clinical applicability and incorporation into standard oncology practice.