Multicenter MRI radiomics targets an elusive breast cancer risk marker
Lymphovascular invasion in breast cancer remains a pathologic finding with outsized clinical weight. It signals that tumor cells have already found a route into vascular or lymphatic channels, a step that can foreshadow metastatic spread, yet definitive confirmation still arrives after surgery. The challenge for surgeons and oncologists is obvious: preoperative imaging rarely captures this biology with enough fidelity to guide decisions before the first incision.
A multicenter retrospective study addressed that gap with dynamic contrast-enhanced magnetic resonance imaging radiomics paired with clinical variables. Investigators analyzed 935 women with pathologically confirmed invasive ductal carcinoma from three institutions, all of whom had preoperative 3.0 T DCE-MRI. The work is notable not because it produced a single best model, but because it compared ten machine-learning classifiers side by side and then asked a harder question: whether any of them stayed reliable outside the hospitals where they were developed.
The answer was mixed, and clinically instructive.
Why lymphovascular invasion is so difficult to predict before the path report
Lymphovascular invasion is not just a histologic checkbox. It reflects a tumor phenotype capable of breaching tissue boundaries, interacting with stromal and endothelial environments, and using vascular channels as a conduit for dissemination. Imaging-based prediction therefore depends on finding subtle spatial and textural patterns that correlate with that invasive behavior, rather than on obvious size alone.
In this study, radiomic feature extraction was extensive. After manual three-dimensional delineation of regions of interest on DCE-MRI images, 1,197 radiomic features were extracted using PyRadiomics after resampling to 2 × 2 × 2 mm³ with B-spline interpolation, intensity normalization, and fixed-bin-width discretization. Feature selection stayed within the training cohort and used a sequence of the Mann-Whitney U test, Pearson correlation pruning with |r| > 0.9, and LASSO regression. That sequencing matters because it limits information leakage and gives the external test sets a more honest chance to reveal overfitting.
Among the ten classifiers, logistic regression and multi-layer perceptron emerged as the most stable across centers. Logistic regression produced AUCs of 0.841 in training, 0.716 in internal validation, 0.632 in test 1, and 0.586 in test 2. The MLP was similarly steady, though the abstract did not provide its full AUC series. In contrast, several ensemble methods performed impressively in the development cohort and then fell apart elsewhere. Gradient Boosting reached 0.991 in training but dropped to 0.696 and 0.532 in external cohorts. XGBoost fell from 0.894 to 0.613 and 0.453.
Preoperative variables helped, but postoperative status changed the picture
The clinical model was intentionally built from variables available before surgery: age, histological grade, Ki-67, MRI axillary lymph node status, and MRI tumor size. A sensitivity version also included postoperative sentinel lymph node status. In multivariable analysis, sentinel lymph node status showed a strong association with lymphovascular invasion, with an odds ratio of 5.743 and a 95% confidence interval of 3.693 to 8.931. MRI tumor size was also independently associated with lymphovascular invasion, with an odds ratio of 1.413 and a 95% confidence interval of 1.186 to 1.684.
The clinical meaning is straightforward. Larger tumors and nodal involvement tend to travel with more aggressive local biology, but sentinel lymph node status is not truly available for preoperative decision-making. The primary clinical model therefore excluded it, preserving the real-world constraint that imaging and routine preoperative data must carry the predictive load alone.
The combined preoperative clinical-radiomics model performed better than the clinical model alone in some settings, but not all. Its AUCs were 0.856 in training, 0.721 in validation, 0.665 in test 1, and 0.603 in test 2. It significantly outperformed the clinical model in training, internal validation, and test 2, with DeLong _p_ values of less than 0.001, 0.013, and less than 0.001, respectively. In test 1, the comparison was not significant (_p_ = 0.306). The model also significantly outperformed the radiomics-only model in test 1 (_p_ = 0.026).
External validation exposed calibration problems that matter at the bedside
The most important caution came from calibration. Even when discrimination remained usable in some cohorts, the combined model’s calibration slopes deteriorated externally to 0.49, 0.28, and 0.21. A slope that low means the predicted probabilities were not well aligned with observed risk in outside institutions. In practical terms, a patient assigned a high predicted probability in one center may not carry the same true risk in another center without recalibration.
That problem is common in radiomics and easy to underestimate. Model performance can look strong inside a development dataset because the classifier learns scanner-specific, protocol-specific, or site-specific patterns along with disease biology. Once the model leaves that environment, the apparent signal can weaken. Here, the training cohort still favored the combined approach, with integrated discrimination improvement showing significant incremental value for radiomics, but that benefit did not persist consistently across external cohorts.
Decision curve analysis and Brier scores were also part of the performance assessment, reinforcing that the investigators did not stop at AUC alone. Still, the broader message stayed the same: radiomics can add signal, but it does not automatically generalize.
What the multicenter data mean for preoperative planning
The study’s most honest contribution is its restraint. Preoperative radiomics can improve discrimination for lymphovascular invasion, and logistic regression or MLP may be better bets than more elaborate ensemble approaches when external robustness is the priority. Yet the gains were not uniform, and the calibration drift across centers means local validation cannot be skipped.
A sensitivity analysis that included postoperative sentinel lymph node status produced higher absolute performance, but that advantage came at the cost of preoperative applicability. In other words, the model performed better when it was allowed to use information that surgeons do not yet have. That tradeoff is not a flaw in the paper; it is the reality of translational imaging research.
For now, the strongest message is methodological. Multicenter benchmarking, external testing, and recalibration are not optional extras in radiomics work. They are the difference between a model that merely looks impressive and one that can plausibly guide care. This dataset makes clear that the gap between internal performance and deployable clinical utility remains wide, especially when the endpoint is as biologically consequential as lymphovascular invasion in invasive ductal carcinoma.