Exempting axillary staging surgery in breast cancer using multimodal ultrasound imaging and radiomics of sentinel lymph nodes.

Background: Sentinel lymph node biopsy (SLNB) is the standard procedure for axillary staging in early-stage breast cancer patients, however, it remains an invasive procedure. The aim of this study is to construct a multicenter, multimodal predictive model based on contrast-enhanced ultrasound (CEUS) and grayscale ultrasound (GSUS) imaging of sentinel lymph nodes (SLNs) in breast cancer patients. The model seeks to preoperatively assess the risk of SLN metastasis in a non-invasive manner, thereby enabling the exemption of unnecessary SLNB for eligible patients. Methods: In this multicenter, multimodal ultrasound radiomics study, eligible breast cancer patients from three medical centers, respectively, the Sichuan Provincial People's Hospital, Yunnan Provincial Cancer Hospital, and Fujian Provincial Cancer Hospital in China, were consecutively enrolled between January 2019 to February 2024, and between February 2024 to July 2024. The enrolled patients had pathologically confirmed breast cancer and underwent CEUS and GSUS imaging of their SLNs. The patients were divided into the following groups: training cohort (n = 763), validation cohort (n = 132), internal independent test cohort (n = 298), prospective internal test cohort 1 (n = 75), prospective external test cohort 2 (n = 51), and prospective external test cohort 3 (n = 55). A deep dual-modal fusion network (DDFN) model was developed to preoperatively predict lymph node metastasis by integrating features from both CEUS and GSUS images of the SLNs. The predictive performance of different models across the test cohorts was evaluated by negative predictive value (NPV), specificity, the area under the ROC curve (AUC), and accuracy. Findings: The DDFN demonstrated superior performance for SLN metastasis prediction compared to single-modality models. In the internal test cohort (n = 298), the DDFN model achieved a NPV of 0.973 (95% CI: 0.956-0.987), which was significantly higher than those of the GSUS model (NPV = 0.941, P = 0.032) and the CEUS model (NPV = 0.958, P = 0.041). The DDFN model also attained the highest AUC of 0.912, significantly outperforming the GSUS model (AUC = 0.782, P = 0.0046) and the CEUS model (AUC = 0.890, P = 0.039). Furthermore, the DDFN model exhibited excellent specificity (0.987), indicating its robustness in accurately distinguishing metastatic and non-metastatic SLNs. This strong performance was consistently maintained across three prospective multicenter test cohorts. The DDFN model yielded NPVs exceeding 0.9 in all cohorts (cohort 1: 0.933; cohort 2: 0.917; cohort 3: 0.909), which were statistically superior to the single-modality models in most comparisons. The AUC values of the DDFN model in the prospective cohorts (0.893, 0.866, and 0.862, respectively) remained high and generally surpassed those of the single-modality approaches. Interpretation: The DDFN model, integrating CEUS and GSUS images, enables preoperative evaluation of SLNs. This method holds promise for assessing axillary lymph nodes (ALNs) preoperatively, identifying patients without SLN metastasis, and exempting them from unnecessary axillary staging surgery. Funding: The study was funded by Key research and development (R&D) projects of Sichuan Science and Technology Department [item 2023YFS0263].