Long-Term Predictive Value of MRI Radiomics in Identifying Patients with Metastasis-Free Breast Cancer Over a 100-Month Follow-Up Period: A retrospective, observational study
Yi-shan Zhao, Yu Ji, Yan-bo Li, Yu-heng Wang, Xiao-xu Ma, Jing-bo Wang, Ping Wang, Hong LuIntroduction:
Postoperative distant metastasis remains a major challenge in breast cancer management, which significantly affects long-term survival among patients. Standard follow-up protocols rely on frequent imaging and clinical evaluations to detect recurrence early. However, such intensive surveillance imposes a financial and psychological burden on both patients and healthcare systems. Recent advances in breast Magnetic Resonance Imaging (MRI)-based radiomics provide new opportunities to capture tumor heterogeneity and may enhance individualized risk prediction. This study aims to assess the long-term predictive value of MRI-based radiomics combined with clinicopathological features in identifying patients with breast cancer and to investigate the potential of this approach in guiding individualized follow-up strategies.
Materials and Methods:
This retrospective study included 388 women with pathologically confirmed breast cancer who underwent preoperative breast MRI at Tianjin Medical University Cancer Institute & Hospital between February 2015 and October 2017. Tumors were segmented on the second post-contrast phase of dynamic contrast-enhanced MRI (DCE-MRI), and radiomic features were extracted using a standardized software configuration. A combined prediction model integrating radiomic features and clinicopathological variables was developed using LASSO-based feature selection and a soft-voting ensemble machine learning framework. Model performance was evaluated in the training and validation cohorts using the area under the Receiver Operating Characteristic (ROC) curve (AUC-ROC) and other classification metrics. In the validation cohort, the optimal cutoff for risk stratification was determined using the Youden index.
Results:
In the validation cohort, the combined model achieved an AUC of 0.686 (95% CI, 0.585–0.788) for predicting postoperative distant metastasis. The optimal cutoff for the predicted probability was 0.362, which stratified patients into high- and low-risk groups with significantly different distant metastasis–free survival (log-rank P < 0.05). SHAP analysis showed that several radiomics texture features and selected clinical variables, including ER status, were major contributors to the model output.
Discussion:
These findings suggest that a combined MRI radiomics–clinical model may provide useful long-term risk stratification for postoperative distant metastasis in breast cancer. SHAP analysis indicated that radiomics texture features and selected clinical variables jointly contributed to the model output. Nevertheless, given the model’s moderate discriminative performance and the supplementary calibration and decision curve results, its absolute risk estimates and potential clinical utility should be interpreted cautiously. At present, the model is best considered an exploratory tool for risk stratification that warrants further validation.
Conclusion:
Radiomics-based modeling using DCE-MRI and clinical data demonstrated potential in effectively predicting long-term metastatic risk in patients with breast cancer and identifying a subset of metastatic-free patients with favorable profiles who may benefit from less intensive surveillance.