Ultrasound-based deep learning-derived radiomics for preoperative prediction of sentinel lymph node metastasis in invasive breast cancer: a two-center retrospective study

Preoperative evaluation of sentinel lymph node (SLN) metastasis facilitates individualized axillary management for patients with invasive breast cancer (IBC). This study aimed to develop and externally validate a deep learning-derived ultrasound radiomics model to predict SLN metastasis. In this retrospective two-center study, 246 pathologically confirmed IBC patients were enrolled. 194 patients from Center A were split into training ( n = 155) and internal test ( n = 39) cohorts, while 52 patients from Center B served as the external validation cohort. Deep features were extracted from preoperative ultrasound images via ResNet50. After feature selection within the training cohort, a gradient boosting decision tree classifier was constructed. Model discrimination was assessed using receiver-operating characteristic analysis, and decision-curve analysis was applied for exploratory net-benefit evaluation. Twenty-two deep learning-derived features were retained. The model achieved areas under the receiver operating characteristic curve (AUC) of 0.734 (95% CI, 0.655–0.813), 0.761 (95% CI, 0.601–0.920), and 0.742 (95% CI, 0.598–0.886) in the training, internal test, and external validation cohorts, respectively. In the external cohort, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were 0.769, 0.650, 0.844, 0.722, and 0.794, respectively. Decision curves were exploratory and do not establish clinical utility. The model showed moderate discrimination for SLN metastasis, with similar AUC point estimates but wide confidence intervals across cohorts; it provides preliminary discrimination and risk ranking only and is not currently clinically actionable. Prospective validation in larger cohorts is required before any clinical use.