AI Reads MRI Scans to Tell Apart Two Breast Cancer Subtypes Without a Biopsy

For decades, the molecular identity of a breast tumor has been locked behind a needle. Surgeons and oncologists rely on biopsy tissue, stained and examined under a microscope, to determine which subtype of breast cancer they are dealing with — and that answer shapes everything from the drugs a patient receives to the intensity of her treatment. Now a team of researchers in China has shown that a machine learning model trained on ordinary MRI scans can make that determination non-invasively, before a single incision is made, and that its conclusions hold up across three independent hospitals. The study, published in BMC Medical Imaging, offers one of the most rigorous demonstrations yet that radiomics — the extraction of vast libraries of quantitative features from medical images — can peer into tumor biology that pathologists currently need tissue to see.

The target of the investigation was a distinction that matters enormously in the clinic but is surprisingly hard to make: the difference between luminal A and luminal B breast cancer. Both subtypes are hormone receptor positive, meaning their cells carry estrogen or progesterone receptors that fuel their growth, and both are generally more treatable than triple-negative or HER2-enriched disease. But they behave differently. Luminal A tumors tend to grow more slowly, carry a better prognosis, and usually respond well to endocrine therapy alone. Luminal B tumors are more aggressive, more likely to proliferate rapidly, and often require chemotherapy in addition to hormone-blocking drugs. The conventional way to separate them is immunohistochemistry, a staining technique that measures receptors and a proliferation marker called Ki-67. The problem is that a biopsy samples only a fragment of the tumor, while molecular characteristics can vary dramatically from one region of a mass to another. A needle that lands in a quiet corner of an aggressive tumor can lead to an underestimated subtype and undertreated disease.

The research team, led by Hui Zhou and Daoyu Yang with corresponding author Xianchun Zeng, set out to capture what the biopsy cannot: the whole tumor. Their approach rests on multiparametric MRI, a protocol that combines three complementary ways of imaging the breast. T2-weighted imaging reveals tissue texture and water content, highlighting edema and necrosis. Diffusion-weighted imaging tracks how freely water molecules move, which reflects cellular density and membrane integrity — densely packed, rapidly dividing cells restrict water diffusion in characteristic ways. Dynamic contrast-enhanced MRI, the workhorse of breast imaging, watches a gadolinium contrast agent wash into and out of the tumor over time, mapping the chaotic vasculature that malignant growths build to feed themselves. Each sequence contributes a different window onto the tumor’s physiology, and each window contains patterns invisible to the human eye.

To convert those patterns into numbers, the researchers used PyRadiomics, an open-source software platform that computes hundreds of quantitative features from a defined region of interest. These features include first-order statistics describing the distribution of pixel intensities, shape metrics capturing the tumor’s geometry and margins, and higher-order texture measures that quantify how intensities are spatially arranged — whether the tumor is homogeneous or patchy, smooth or irregular at scales ranging from a few pixels to the entire lesion. A single MRI study can yield well over a thousand such features, far too many to feed directly into a classifier without the model simply memorizing the training data. The team therefore built a four-stage filtering pipeline: variance thresholding to discard features that barely varied across patients, Pearson correlation analysis to remove redundant measures that duplicated one another, an ANOVA F-test to keep only features statistically associated with the subtype labels, and finally L1-regularized logistic regression, a technique that shrinks unimportant feature weights to zero, with five-fold cross-validation to guard against overfitting.

One further statistical hurdle deserved special attention. In the study cohort of 130 patients with pathologically confirmed invasive ductal carcinoma, luminal B tumors outnumbered luminal A tumors 86 to 44 — a ratio that reflects real-world prevalence but that would bias a naive classifier toward always predicting the more common class. The researchers applied the Synthetic Minority Over-sampling Technique, or SMOTE, which generates synthetic examples of the underrepresented luminal A class by interpolating between existing cases in the high-dimensional feature space. Crucially, they applied SMOTE only within the training folds of the cross-validation procedure, ensuring that no synthetic data leaked into the validation sets and inflated performance estimates. It is the kind of methodological discipline that separates credible machine learning studies from the many that quietly overstate their accuracy.

With the feature set refined and the class balance corrected, the team compared nine machine learning classifiers head to head, ranging from classical algorithms like logistic regression and support vector machines to gradient-boosted decision tree ensembles. The winner was XGBoost, an algorithm that builds a cascade of simple decision trees, each one trained to correct the residual errors of its predecessors. XGBoost has become a favorite in radiomics research because it handles heterogeneous feature types gracefully, resists overfitting through built-in regularization, and consistently tops machine learning competitions on tabular data of exactly this kind. The final model fused the radiomics features with clinical variables, and its performance was assessed with the area under the receiver operating characteristic curve, or AUC, a metric that captures how well the model separates the two classes across all possible decision thresholds.

The results were strongest at the institution where the model was developed. Using data from Center 1, comprising 64 patients, the clinical-radiomics fusion model achieved an AUC of 0.8944, with a standard deviation of 0.0711 across the cross-validation folds — a level of discrimination that approaches the reliability of the immunohistochemistry it is meant to complement. External validation, the true test of any predictive model, followed at two hospitals the model had never seen. At Center 2, with 36 patients, the AUC was 0.8483, still comfortably in the range considered clinically useful. At Center 3, with 30 patients, performance declined to 0.7269, a drop the authors attribute to the smaller sample size and the inevitable differences in scanner hardware, imaging protocols, and patient populations between institutions. Even so, the model retained meaningful discriminatory power at all three sites, and pairwise DeLong tests confirmed that the multiparametric fusion model significantly outperformed models built on any single MRI sequence alone. Decision curve analysis, which quantifies the net clinical benefit of acting on a model’s predictions at various threshold probabilities, showed the fusion model offered superior net benefit compared with simpler alternatives.

Perhaps the most forward-looking element of the study is its commitment to transparency. Black-box predictions are a hard sell in oncology, where clinicians need to understand why a model reached its conclusion before trusting it with treatment decisions. To open the box, the researchers applied SHAP analysis, a technique rooted in cooperative game theory that assigns each input feature a contribution value for every individual prediction. The analysis revealed that features derived from diffusion-weighted imaging and dynamic contrast-enhanced MRI were the most influential predictors of subtype, and — critically — the importance rankings remained consistent across all three centers. That consistency suggests the model is latching onto genuine biological signal, such as the cellularity captured by diffusion imaging and the vascular permeability captured by contrast dynamics, rather than artifacts of a particular scanner or protocol. It also gives radiologists a concrete hypothesis about which imaging phenomena encode proliferation status, the biological variable that most cleanly separates luminal A from luminal B disease.

The authors are careful to position the model as a complement to immunohistochemistry rather than a replacement. Biopsy remains the gold standard, and no regulatory agency is likely to clear an imaging classifier to stand alone in subtype determination. But the model addresses a genuine weakness of the current standard: sampling error. Because radiomics features are computed across the entire tumor volume, they integrate information from every region of the lesion, capturing intratumoral heterogeneity that a core needle can miss. In practice, a model like this could serve as a second opinion at the time of diagnostic MRI, flagging cases where the imaging signature of the tumor disagrees with the biopsy result and prompting clinicians to consider repeat sampling or more aggressive treatment planning. It could also help stratify patients in regions or healthcare systems where immunohistochemistry panels are delayed or incomplete.

The study is not without limitations, and the authors acknowledge them. The cohort of 130 patients is modest by the standards of deep learning, and the luminal A group in particular is small, which widens the confidence intervals around the performance estimates. The retrospective design means the model was tested on patients whose scans were acquired under each institution’s routine protocols rather than a standardized imaging regimen, which cuts both ways: it makes the external validation results more realistic, but it also leaves open the question of how much performance could improve with harmonized acquisition. The decline in AUC at Center 3 is a reminder that generalization remains the central challenge of clinical machine learning. Still, the trajectory of the field is clear. As imaging datasets grow and multicenter collaborations mature, quantitative MRI combined with interpretable machine learning is moving steadily from the research lab toward the radiology reading room — and for the hundreds of thousands of women diagnosed with hormone receptor positive breast cancer each year, the prospect of a fuller, faster, and less invasive picture of their tumor’s biology is a development worth watching closely.

Subject of Research: Multiparametric MRI radiomics and machine learning for preoperative differentiation of luminal A and luminal B breast cancer subtypes

Article Title: Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter study

Article References: Multiparametric MRI-based radiomics combined with machine learning for preoperative differentiation of luminal A and luminal B breast cancer: a multicenter study. (n.d.). https://doi.org/10.1186/s12880-026-02805-2

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02805-2

Keywords: breast cancer, radiomics, multiparametric MRI, machine learning, XGBoost, luminal A, luminal B, DCE-MRI, diffusion-weighted imaging, SHAP, multicenter validation, tumor heterogeneity