Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer Using a Three-tier MRI Radiomics Model
Hong Li1, Weiqing Huang1, Jiefeng Liang1
1Department of Radiology, The Second Affiliated Hospital, Guangzhou Medical University, Guangzhou, Guangdong, China (H.L., W.H., J.L., S.H., H.C.).
Rationale And Objectives:
To develop and validate a three-tier model incorporating clinical, qualitative magnetic resonance imaging (MRI), and radiomics features for preoperative prediction of axillary lymph node (ALN) metastasis in breast cancer, using nested cross-validation to ensure unbiased performance estimates.
Materials And Methods:
This retrospective study included 494 patients with pathologically confirmed breast cancer who underwent preoperative MRI (dynamic contrast-enhanced [DCE] and T2FS-STIR sequences) between July 2018 and August 2024. Three progressive models were developed as follows: Model 1 (clinical variables: age, tumor size, estrogen receptor, progesterone receptor, and human epidermal growth factor receptor 2), Model 2 (Model 1 + qualitative MRI features: peritumoral edema, time-intensity curve pattern, multifocality, field strength), and Model 3 (Model 2 + radiomics scores from DCE and T2FS sequences). Radiomics features were extracted using PyRadiomics and selected through variance-correlation-univariate-LASSO filtering. True nested five-fold cross-validation repeated three times ensured that feature selection and model training were performed independently within each training fold. Model performance was compared using Delong test with Bonferroni correction. SHAP analysis provided model interpretability.
Results:
Model 3 achieved significantly higher AUC (0.769, 95% CI: 0.724-0.808) compared to Model 1 (0.676, 95% CI: 0.626-0.722; ΔAUC = +0.093, p<0.001) and Model 2 (0.735, 95% CI: 0.686-0.776; ΔAUC = +0.034, p = 0.002). All comparisons remained significant after Bonferroni correction (α = 0.017). Feature selection demonstrated moderate stability, with 26.8 ± 4.0 (range: 16-32) DCE and 30.8 ± 4.3 (range: 25-38) T2FS features selected per fold. SHAP analysis revealed T2FS-derived radiomics (mean |SHAP| = 0.811) and DCE-derived radiomics (mean |SHAP| = 0.492) as the most important predictors. Bootstrap validation confirmed model stability (optimism = +0.001). The model showed good calibration (Brier score = 0.201). Decision curve analysis demonstrated clinical utility across threshold probabilities 0.20-0.60. Risk stratification achieved negative predictive value of 71.5% (95% CI: 63.2%-78.8%) for low-risk and positive predictive value of 80.2% (95% CI: 73.9%-85.4%) for high-risk groups.
Conclusion:
The three-tier MRI radiomics model significantly improves preoperative ALN metastasis prediction. The nested cross-validation approach ensures credible performance estimates for potential clinical implementation.
