Digital breast tomosynthesis-based radiomics for prediction of prognosis in breast cancer: a multicenter study.

Background: Breast cancer threatens women's health, and predicting its prognosis facilitates early therapeutic intervention. This study aims to develop radiomics models and combined models based on digital breast tomosynthesis (DBT) for predicting breast cancer prognosis and conducting interpretability analysis. Methods: Patients pathologically diagnosed with invasive breast cancer at Fudan University Shanghai Cancer Center from January 2019 to August 2020 were retrospectively included and randomly divided into a training set and a testing set at a 7:3 ratio. An independent external validation set was constructed using invasive breast cancer patients who visited Ruijin Hospital and The Affiliated Hospital of Qingdao University from December 2021 to August 2022. Disease-free survival (DFS) served as the endpoint. Univariate and multivariate Cox regression analyses were performed to identify prognosis-associated conventional imaging features on DBT. Radiomics features were extracted from the maximum layer of lesions in the craniocaudal (CC) and mediolateral oblique (MLO) views of DBT images. Selected radiomics features were incorporated into the Cox proportional hazards model to predict prognosis and a combined model in conjunction with conventional imaging features was constructed. Stratified assessment was conducted for evaluating the model performance by comparing the C-index value, the area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), and calibration curves. Nomograms and Kaplan-Meier curves were plotted to stratify the disease risks. Additionally, SHapley Additive exPlanations (SHAP) were employed to carry out the interpretability analysis. Results: A total of 395 patients were enrolled in the training and testing cohorts, whereas the validation cohort had 140 patients. High-density masses (P=0.01) and axillary adenopathy (P<0.001) were identified as independent factors associated with DFS. Eight radiomics features were ultimately incorporated into the model. In the validation set, the radiomics model exhibited the C-index value of 0.71, while that of the combined model was 0.76. Based on the combined model for stratified prediction, the AUC values for predicting 1-, 2-, and 5-year DFS in the testing set were 0.73, 0.74, and 0.76. In the validation set, the AUC values for predicting 1- and 2-year DFS were 0.74 and 0.76. Both DCA curves and calibration curves confirmed the clinical utility of the combined model. Kaplan-Meier curves showed that the combined model stratified patients into high-risk and low-risk groups (P values were <0.001 in the training set, 0.03 in the testing set, and 0.03 in the external validation set). SHAP analysis revealed that radiomics features derived from wavelet transformation and those from the CC view contributed more substantially and carried higher weights among the selected features. Conclusions: Radiomics based on DBT have potential to predict breast cancer prognosis in terms of short-term DFS, with the combined model exhibiting superior efficacy. SHAP analysis is conducive to mining imaging biomarkers related to prognosis.