LVI describes the presence of tumor cell clusters inside endothelial-lined lymphatic vessels or blood vessels surrounding an invasive breast cancer. It is an independent predictor of local recurrence, distant metastasis, and poor prognosis, yet confirming it today depends entirely on postoperative pathology. Surgeons learn whether tumor cells had already entered vessels only after the tumor has been removed, which limits their ability to fine-tune the extent of axillary lymph node dissection or to personalize adjuvant treatment decisions. A reliable preoperative, noninvasive estimate of LVI status has therefore been a long-standing unmet need in breast oncology.

The new study tackles that gap using dynamic contrast-enhanced magnetic resonance imaging, or DCE-MRI, already a core preoperative modality thanks to its exceptional soft-tissue resolution and its sensitivity to tumor vascularization. The researchers combined DCE-MRI with radiomics, a technique that extracts large numbers of quantitative image features invisible to the human eye, and with machine learning, allowing algorithms to decode tumor heterogeneity from routine scans. While earlier efforts have used mammography-based tomosynthesis or single-center MRI cohorts, systematic comparisons of many algorithms on multicenter DCE-MRI data, with rigorous external validation, remained scarce.

Retrospectively, the team assembled 912 female patients aged 23 to 81 years from two independent medical centers, Guangdong Maternal and Child Health Hospital and The First Affiliated Hospital of Jinan University, all scanned with a Philips 3.0T scanner within two weeks before surgery. Data from 757 patients at the first center formed the model development set, split 7:3 into training and internal validation subsets using stratified sampling on LVI status. Data from 155 patients at the second center served as a completely unseen external test set, a design that directly probes how well the model generalizes to a different hospital population rather than merely memorizing patterns from its own data.

The radiomics pipeline began with manual, blinded contouring of the entire tumor on the arterial phase of the six-phase contrast sequence, producing a three-dimensional volume of interest reviewed by senior radiologists. Using PyRadiomics, the team extracted 1,197 features per patient, spanning 234 histogram, 14 morphological, 884 texture, and 65 higher-order descriptors. A multistep reduction process followed: correlation filtering removed redundant features, univariate analyses retained only those significantly associated with LVI, and LASSO regression with 10-fold cross-validation ultimately selected 18 features with the highest predictive value for modeling.

Crucially, the researchers did not commit to a single algorithm. They built and systematically compared 10 machine learning models, including logistic regression, support vector machine, K-nearest neighbors, random forest, ExtraTrees, XGBoost, LightGBM, gradient boosting, AdaBoost, and multilayer perceptron, tuning each via grid search and comparing them with ROC analysis and the DeLong test. The ExtraTrees model emerged as the best radiomics performer, achieving an AUC of 0.653 and accuracy of 0.640 on the validation set, a deliberately conservative benchmark compared with many single-center studies that report higher numbers without independent external testing.

In parallel, multivariate logistic regression identified three independent clinical risk factors for LVI: sentinel lymph node status, estrogen receptor status, and the time-intensity curve pattern of contrast enhancement. The team then integrated the ExtraTrees radiomics score with these clinical predictors to construct a combined model, visualized as a practical nomogram for individualized bedside estimation of LVI probability. The combined model outperformed either the radiomics or clinical model alone across all cohorts, reaching an AUC of 0.796 on the training set, 0.704 on internal validation, and 0.703 on the external test set, with calibration curves and decision curve analysis confirming good agreement and clinical net benefit across a broad range of threshold probabilities.

The authors argue that their relatively modest external AUC values are not a weakness but a feature of honest validation. Many published radiomics studies reporting AUCs above 0.80 relied on single-center data, lenient validation schemes, or highly homogeneous patient subsets. By enrolling all molecular subtypes, including luminal A, luminal B, HER2-overexpressing, and triple-negative cancers, and by testing on a genuinely independent center with real-world data heterogeneity, the study offers a more realistic assessment of how such models behave outside their birthplace. The researchers also note that no dedicated image harmonization strategy, such as ComBat, was applied before feature extraction, which may partly explain the performance drop on external data and marks a target for future improvement.

Limitations remain. The retrospective design carries inherent selection bias, the external cohort was relatively small, and tumor delineation still depends on manual contouring by experienced radiologists, a process that is time-consuming and subject to inter-observer variability. The two centers also used the same scanner vendor, and prospective validation in larger, more diverse populations will be essential before routine clinical deployment. The team outlines next steps that include incorporating diffusion-weighted MRI parameters, applying explainable AI to reveal the imaging basis of model decisions, and integrating the tool into clinical decision support systems to measure its actual impact on surgical planning and patient outcomes.

Even with those caveats, the study delivers a clinically meaningful proof of concept: a noninvasive, preoperative estimate of lymphovascular invasion is achievable when carefully selected radiomic features from DCE-MRI are combined with independent clinical risk factors and validated the hard way. For patients, that could mean smarter surgical decisions and more personalized adjuvant therapy planned before the operation rather than pieced together after it. For the growing field of radiomics, it is a reminder that multicenter external validation, systematic algorithm comparison, and honest reporting of performance are what separate a promising research model from one ready for the clinic.

Subject of Research: Preoperative prediction of lymphovascular invasion in breast cancer using DCE-MRI radiomics and machine learning

Article Title: Machine Learning Comparison and Combined Model Optimization Based on DCE‐MRI Radiomics for Preoperative Assessment of Lymphovascular Invasion in Breast Cancer: A Multicenter Study

Article References: Li, H., Xiao, Q., Zeng, Y., Wang, X., & Zhang, Y. (2026). Machine Learning Comparison and Combined Model Optimization Based on DCE ‐ MRI Radiomics for Preoperative Assessment of Lymphovascular Invasion in Breast Cancer: A Multicenter Study. Cancer Reports, 9(9), Article e70678. https://doi.org/10.1002/cnr2.70678

Image Credits: AI Generated

DOI: 10.1002/cnr2.70678

Keywords: breast cancer, lymphovascular invasion, DCE-MRI, radiomics, machine learning, ExtraTrees, nomogram, multicenter study, external validation, clinical risk factors, LASSO regression, preoperative assessment