Interpretable machine learning model based on intratumoral and peritumoral MRI for predicting tumor budding grade in breast cancer

Objective Tumor budding (TB) is a histopathological indicator associated with aggressive tumor behavior and poor prognosis in breast cancer. However, TB assessment is mainly dependent on postoperative pathological examination, limiting its availability for preoperative risk stratification and treatment decision-making. This study aimed to develop a non-invasive nomogram integrating intratumoral and peritumoral radiomic features with clinicopathological variables to predict the tumor budding grade of breast cancer. Methods Clinicopathological data of patients with pathologically confirmed invasive breast cancer were retrospectively analyzed, and their preoperative magnetic resonance imaging (MRI) scans were collected. Volume of interests (VOIs) of the tumors were manually delineated on the dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and diffusion-weighted imaging (DWI) sequences. Peritumoral VOIs were obtained by expanding the intratumoral VOIs outward by 3 mm. Radiomic features were extracted and screened from the intratumoral, peritumoral VOIs separately, and six prediction models were established. A nomogram was further constructed by incorporating clinicopathological variables. The performance of all models was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence interval (95% CI). Calibration curves and decision curve analysis (DCA) were applied to evaluate the clinical applicability of the predictive models, and SHAP plots were used for further visual interpretation of the model results. Results A total of 281 patients with breast cancer were enrolled in this study, including 175 cases in the training set, 76 cases in the internal validation set and 30 cases in the external validation set. Among the clinicopathological variables, Ki-67 expression and enhancement pattern were significantly correlated with tumor budding grade, and the established clinicopathological model exhibited good predictive performance. The combined intratumoral-peritumoral model based on fused DCE-MRI and DWI features outperformed all other single models, with AUC values of 0.854, 0.760 and 0.755 in the training, internal validation and external validation sets, respectively. The nomogram model achieved even better predictive efficacy, with AUC values of 0.906, 0.845 and 0.825 in the training, internal validation and external validation sets, respectively. Conclusion The proposed MRI-based nomogram provides a reliable and interpretable approach for non-invasive preoperative prediction of tumor budding grade in breast cancer. By enabling early identification of patients with aggressive tumor characteristics before surgery, this model may facilitate individualized risk stratification and treatment planning, and represents a potential complementary tool to conventional pathological assessment. Further multicenter studies are warranted to validate its clinical applicability and promote its integration into personalized breast cancer management.