Enhancing the Prediction of Lymph Node Metastasis in Early Breast Cancer Using Deep Learning on Routine Full-Breast Mammograms

Background: With a trend toward de-escalation of axillary surgery in breast cancer, prediction models incorporating imaging modalities can help reassess the need for surgical axillary staging. Although mammography is routinely performed for breast cancer imaging, its potential in nodal staging remains underutilized. This study aims to employ advancements in deep learning (DL) to comprehensively evaluate the potential of routine mammograms for predicting lymph node metastasis (LNM) in preoperative clinical settings. Methods: This retrospective study included 1,265 cN0 T1-T2 breast cancer patients, comprising 368 node-positive and 897 node-negative cases, diagnosed from 2009-2017 at three Swedish institutions. Patients diagnosed in 2017 were assigned to the independent test set (n=123, site 2) and the external test set (n=103, site 3), while the remaining patients (n=1,039, site 1 and 2) were used for model development and double cross-validation. A neck module, in conjunction with a ResNet backbone pretrained on unlabeled mammograms, was developed to extract global information from full-breast or region-of-interest (ROI) mammograms by predicting five cancer outcomes. Clinicopathological characteristics were combined with the learned mammogram features to predict LNM collaboratively. The models were evaluated using area under the receiver operating characteristic (ROC) curve (AUC), calibration, and decision curve analysis. Results: Compared to models using only clinical variables, incorporating full-breast mammograms with preoperative clinical variables improved the ROC AUC from 0.690 ± 0.063 (SD) to 0.774 ± 0.057 in the independent test set and from 0.584 ± 0.068 to 0.637 ± 0.063 in the external test set. The combined model showed good calibration and, at sensitivity ≥ 90%, achieved a better net benefit, and a higher sentinel lymph node biopsy reduction rate of 41.7% in the independent test set. Full-breast mammograms showed comparable ability to tumor ROIs in predicting LNM. Conclusion: Our findings underscore that routine mammograms, particularly full-breast images, can enhance preoperative nodal status prediction. They may substitute key postoperative predictors such as pathological tumor size and multifocality, aiding patient stratification before surgery. Interestingly, the added predictive value of mammography was consistent across all sites, whereas the overall performance varied over time periods and sites, likely due to advancements in equipment and procedures.