A deep learning radiomics model for predicting non-sentinel lymph node metastases in early-stage breast cancer patients

To develop and validate a deep learning radiomics model to predict non-sentinel lymph node (NSLN) metastases in early-stage breast cancer patients with 1–2 positive sentinel lymph node (SLN) metastases. This retrospective and prospective study encompassed 1,647 patients. Clinical, pathological information, and axillary ultrasound (AUS) findings, collected. Radiomic features of breast cancer lesions were extracted from the ultrasound images. We developed predictive models based on clinical factors alone (C model), clinical factors coupled with AUS (CA model), and clinical factors integrated with both AUS and radiomic features (CAR model). The predictive performance of each model was evaluated via the area under the curve (AUC), decision curve analysis (DCA), and calibration curve analysis. The AUC values for the C model, CA model and CAR model in the test cohort were 0.812, 0.850, and 0.994, respectively. Notably, the CAR model exhibited significantly superior predictive capability compared to both the C model and CA model. In subgroups analyses, the CAR model also achieved the optimal predictive performance. The DCA curve confirmed that the CAR model possessed significant clinical implications. The CAR model had the capability to predict NSLN metastases in early-stage breast cancer with 1–2 positive SLN metastases. Axillary lymph node dissection is controversial in patients with 1–2 positive sentinel lymph node metastases because non-sentinel lymph node is negative in some of these patients. In order to identify this group patients, this study built an algorithm with a quite large group of patients. This algorithm built with all valuable informantion of patients including clinical and pathological information, axillary ultrasound findings, as well as ultrasound radiomic features of breast cancer lesions. This algorithm is able to identify above-mentioned patients who would not benefit from axillary lymph node dissection with area under the curve as high as 0.994. The performance of our algorithm is superior to existing models such as MSKCC, Tenon and is robust across different subgroups.