Tumor budding, the scattering of individual tumor cells or small clusters at the invasive edge of a breast tumor, is a histopathological sign linked to aggressive tumor behavior and poor prognosis. In a new study published in BMC Medical Imaging, researchers report that an interpretable machine learning nomogram built from preoperative MRI scans can predict the tumor budding grade of breast cancer non-invasively, before surgery, achieving an area under the receiver operating characteristic curve of 0.906 in the training set, 0.845 in internal validation and 0.825 in external validation.
The motivation is practical. Tumor budding assessment currently depends on postoperative pathological examination of surgically removed tissue, which limits its availability for preoperative risk stratification and treatment decision-making. The 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, giving clinicians a potential window into this aggressive feature while treatment plans are still being formed.
The retrospective analysis enrolled 281 patients with pathologically confirmed invasive breast cancer, divided into a training set of 175 cases, an internal validation set of 76 cases and an external validation set of 30 cases. Preoperative MRI scans were collected for each patient, and volumes of interest covering the tumors were manually delineated on both the dynamic contrast-enhanced MRI and diffusion-weighted imaging sequences. Peritumoral volumes were then generated by expanding each intratumoral volume outward by 3 millimeters, capturing the tissue immediately surrounding the tumor as well as the tumor itself.
Radiomic features were extracted and screened separately from the intratumoral and peritumoral volumes, and six prediction models were established from these feature sets. The combined intratumoral-peritumoral model, based on fused DCE-MRI and DWI features, outperformed all other single models, with AUC values of 0.854 in the training set, 0.760 in internal validation and 0.755 in external validation. The fusion of features from inside and around the tumor proved more informative than either region alone, consistent with the idea that the tumor and its microenvironment carry complementary imaging signals.
Clinicopathological variables contributed their own predictive signal. Among the variables examined, Ki-67 expression and the enhancement pattern seen on MRI were significantly correlated with tumor budding grade, and a clinicopathological model built on these variables showed good predictive performance. The final nomogram incorporated these clinicopathological variables together with the radiomic features, and this combined model achieved the strongest results of the study, with AUC values of 0.906, 0.845 and 0.825 in the training, internal validation and external validation sets respectively.
Model performance was assessed using the area under the receiver operating characteristic curve with 95% confidence intervals, and calibration curves together with decision curve analysis were applied to evaluate the clinical applicability of the predictive models. These analyses address a question that AUC values alone cannot: whether the predicted probabilities track observed outcomes closely enough, and whether acting on the model’s predictions would yield a net benefit compared with alternative decision strategies across a clinically relevant range of thresholds.
A distinguishing feature of the work is its emphasis on interpretability. Rather than presenting a black-box classifier, the researchers used SHAP plots for further visual interpretation of the model results. SHAP, which stands for SHapley Additive exPlanations, is a technique from game theory that quantifies how much each input feature pushes an individual prediction toward or away from a given outcome. In a clinical setting, where radiologists and oncologists need to understand why a model flags a tumor as likely high-grade, such visual explanations can make the difference between an accepted decision-support tool and an ignored one.
The clinical logic behind the effort is straightforward. If MRI-based prediction can identify patients with aggressive tumor characteristics before surgery, the model may facilitate individualized risk stratification and treatment planning, and it represents a potential complementary tool to conventional pathological assessment. The authors position it as an addition to, not a replacement for, the microscopic examination that remains the reference standard for tumor budding grading.
Several limitations temper the enthusiasm. The study was retrospective, the external validation set contained only 30 patients, and the authors themselves state that further multicenter studies are warranted to validate the model’s clinical applicability and promote its integration into personalized breast cancer management. Manual delineation of tumor volumes is also labor-intensive, and the model’s performance in a prospective, real-world workflow remains to be demonstrated.
Even so, the study adds to a growing body of work showing that radiomics, the high-throughput extraction of quantitative features from medical images, can surface biological information that images were not traditionally thought to contain. By fusing signals from the tumor interior and its 3-millimeter peritumoral rim across two MRI sequences, and by wrapping the result in an interpretable nomogram, the researchers have produced a tool that predicts a prognostically important pathological grade before the first incision is made. Whether it earns a place in routine practice will depend on the multicenter validation the authors call for, but the reported performance marks a measurable step toward preoperative, image-based assessment of tumor aggressiveness in breast cancer.
Subject of Research: Machine learning on intratumoral and peritumoral MRI features to predict tumor budding grade in breast cancer
Article Title: Interpretable machine learning model based on intratumoral and peritumoral MRI for predicting tumor budding grade in breast cancer
Article References: Sun, M., Zhao, W., Li, R., Wang, Y., Geng, Z., Zhang, H., Liu, F., Zhang, L., Zhou, Y., Sui, B., Lu, S., & Dong, H. (2026). Interpretable machine learning model based on intratumoral and peritumoral MRI for predicting tumor budding grade in breast cancer. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02885-0
Image Credits: AI Generated
DOI: 10.1186/s12880-026-02885-0
Keywords: breast cancer, tumor budding, MRI, radiomics, machine learning, nomogram, DCE-MRI, DWI, SHAP, Ki-67, preoperative prediction, BMC Medical Imaging
News Source: Nathaniel Bowman. (November 8, 2026). MRI-based machine learning model predicts aggressive tumor budding in breast cancer before surgery. Scienmag.



