Abstract P2-09-11: Improved Prediction of Axillary Lymph Node Metastasis in Early-Stage Breast Cancer Using Deep Learning on Routine Mammography
Daqu Zhang, Looket Dihge, Ida Arvidsson, Pär‐Ola Bendahl, Magnus Dustler, Julia Ellbrant, Kim Gulis, Malin Hjärtström, Mattias Ohlsson, Cornelia Rejmer, David Maria Schmidt, Sophia Zachrisson, Patrik Edén, Lisa Rydén · Clinical Cancer Research · 2025
Abstract Background: Sentinel lymph node biopsy is the standard axillary nodal staging procedure and is routinely performed. With a trend towards de-escalation of axillary surgery, recent studies indicate that prediction models incorporating imaging modalities can reassess the necessity of surgical axillary nodal staging. However, the clinical utility of MRI is constrained by accessibility, and ultrasound is highly operator-dependent. Mammography, the primary imaging modality for all breast cancer patients, has drawn little attention in nodal staging. This study aims to employ advancements in deep learning (DL) to comprehensively evaluate the potential of routine mammography for predicting nodal metastasis in preoperative clinical settings. Methods: The study included 1,281 breast cancer patients (age: 62.7 ± 11.5 years, tumor size: 15.3 ± 8.0 mm) diagnosed between 2009 and 2017 at two hospitals in Region Skåne, Sweden. Among these patients, 378 were node-positive and 903 were node-negative. Patients diagnosed in 2017 (n=126) were assigned to the test set, while those from 2009-2016 (n=1,155) were used for model development and cross-validation by period and region (2009-2012 at site 1, 2015-2016 at site 2). Input characteristics included routine preoperative clinical data and features from core needle biopsies: age, BMI, menstrual status, mode of detection, histological grade, histopathological type, and molecular subtype. DL models were constructed in two steps. First, a vision transformer was developed to learn task-specific mammographic features through supervised learning, predicting tumor size, multifocality, lymphovascular invasion, and lymph node metastasis, using two resolutions: the region of interest (ROI) emphasizing the tumor and the full mammographic image. Next, prediction models were trained using both clinical and mammographic features. Results: Double cross-validation on the development set showed that models using only preoperative clinical variables achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.629 ± 0.020. Incorporating ROI-based mammographic features increased the AUC to 0.670 ± 0.018, while utilizing full images resulted in a comparable AUC of 0.669 ± 0.019. Predictive accuracy was improved by 3.4% for patients from 2009 to 2012 at site 1 and by 11.9% for patients from 2015 to 2016 at site 2. Notably, on the independent test set from 2017 at site 2, the combined model of preoperative clinical variables and full mammograms achieved an AUC of 0.784 ± 0.044, outperforming the clinical model that used postoperative tumor size and multifocality (AUC of 0.731 ± 0.051). Thus, DL applied to routine mammography enhanced the prediction of nodal metastasis by 10.6% compared to preoperative clinical models. Conclusion: Our findings underscore that routine mammograms, particularly full images, can enhance nodal status prediction in clinical models. They have the potential to compensate for key postoperative predictors such as tumor size and multifocality, aiding in patient stratification prior to surgery. Interestingly, the added value of mammography for nodal staging varied largely across different periods and regions due to advancements in screening equipment and procedures. Citation Format: Daqu Zhang, Looket Dihge, Ida Arvidsson, Pär-Ola Bendahl, Magnus Dustler, Julia Ellbrant, Kim Gulis, Malin Hjärtström, Mattias Ohlsson, Cornelia Rejmer, David Schmidt, Sophia Zachrisson, Patrik Edén, Lisa Rydén. Improved Prediction of Axillary Lymph Node Metastasis in Early-Stage Breast Cancer Using Deep Learning on Routine Mammography [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P2-09-11.