MRI Radiomics Model Predicts Breast Cancer Chemotherapy Response by Reading the Tumor and Its Surroundings

One of the most consequential decisions in the treatment of stage II and III breast cancer is made before a single dose of chemotherapy is given. Neoadjuvant chemotherapy, in which drugs are administered before surgery to shrink tumors and improve surgical outcomes, works remarkably well for some patients and disappointingly for others. When it works completely, the tumor vanishes under the microscope, a result pathologists call pathological complete response, or pCR. That outcome is strongly linked to better long-term survival. The problem is that predicting, at the start of treatment, which patients will achieve it remains one of the stubborn challenges of breast oncology, because breast tumors are anything but uniform in their biology and behavior.

A new study published in BMC Medical Imaging by Chao Zheng, Bo Wang and colleagues at Hanzhong Central Hospital and collaborating institutions in China takes aim at that prediction problem using an approach that has been gaining momentum across oncology: radiomics. Radiomics is a computational technique that converts the pixel-level information hidden inside standard medical images into large sets of quantitative features. Where a radiologist sees a tumor’s shape and enhancement pattern, a radiomics pipeline extracts hundreds of numerical descriptors capturing texture, intensity distribution, spatial relationships and fine-grained heterogeneity that the human eye cannot resolve. The premise is that these image-derived signatures act as a non-invasive window into the tumor’s underlying biology, including the chaotic architecture of its microenvironment.

The research team assembled a training dataset of 351 patients with stage II or III breast cancer, each with complete magnetic resonance imaging data and electronic health records. From the MRI scans they built two separate machine learning models distinguished by a single, deceptively simple design choice: what region of the image the algorithm was allowed to study. The first model analyzed only the tumor itself, the conventional approach in most radiomics studies. The second analyzed the tumor together with the peritumoral region, the rim of tissue immediately surrounding the tumor. That surrounding zone, often dismissed as unremarkable on visual inspection, is in fact a biologically active frontier where tumor cells interact with immune infiltrates, stromal cells and vasculature, and it is precisely where tumor microenvironment heterogeneity leaves its imaging fingerprints.

The results were unambiguous. The model that incorporated both the tumor and its peritumoral surroundings achieved an area under the curve, or AUC, of 0.71, while the tumor-only model reached only 0.60. DeLong’s test, a standard statistical procedure for comparing the performance of two correlated receiver operating characteristic curves, confirmed that the difference was statistically significant, with a p-value of 0.023. In practical terms, the AUC measures a model’s ability to distinguish patients who will achieve pCR from those who will not, with 0.5 representing random guessing and 1.0 representing perfect discrimination. The eleven-point improvement may sound modest, but in a prediction task as difficult as treatment response, and achieved purely by widening the algorithm’s field of view, it carries real scientific weight.

The study went beyond aggregate model performance to ask which specific imaging features drove the predictions. Global feature importance analysis identified three dominant factors: hormone receptor status, a molecular characteristic of the tumor itself; and two radiomic features with technical names that encode meaningful information about image texture. The first, IMC1, or Informational Measure of Correlation, is a gray-level run-based feature that quantifies the degree of spatial correlation and organization among pixel intensities. In essence, it measures how orderly or disordered the texture of the tumor region appears on MRI. The second, LDHGLE, or Large Dependence High Gray Level Emphasis, captures the prevalence of large, contiguous regions of high signal intensity, reflecting patterns of enhancement that relate to tissue density, vascularity and cellularity.

Perhaps the most clinically intriguing findings emerged when the researchers stratified patients by molecular subtype. Among patients with triple-negative breast cancer, the aggressive subtype that lacks hormone receptors and HER2 expression and has historically had the fewest targeted treatment options, those whose tumors showed an IMC1 value below −0.188 were significantly more likely to achieve pCR after neoadjuvant chemotherapy, with a p-value of 0.044. A parallel pattern appeared in HER2-positive patients, where an IMC1 value below −0.247 was associated with a higher likelihood of complete pathological response, with a p-value of 0.046. These thresholds suggest that a single quantitative descriptor of tumor texture, readable from a routine MRI scan, may carry different predictive information in different molecular contexts, echoing the well-established principle that breast cancer’s subtypes respond to chemotherapy through distinct biological routes.

What makes this result conceptually important is the link it forges between imaging and tumor microenvironment heterogeneity. The tumor microenvironment, the ecosystem of immune cells, fibroblasts, blood vessels and extracellular matrix that surrounds malignant cells, is now recognized as a decisive factor in whether chemotherapy and immunotherapy succeed. But characterizing that environment traditionally requires biopsy and tissue analysis, which sample only a fraction of a heterogeneous tumor and cannot be repeated easily over time. Radiomics offers a complementary strategy: because MRI captures the aggregate effect of tissue composition across the entire tumor and its margins, texture features such as IMC1 and LDHGLE may serve as indirect, spatially comprehensive proxies for microenvironmental variation that a needle biopsy would miss.

The study’s design also reflects a growing recognition in the radiomics field that where you measure matters as much as what you measure. Defining the region of interest is one of the most consequential steps in any radiomics pipeline, and most published studies default to the tumor boundary alone. By explicitly comparing tumor-only and tumor-plus-peritumoral models on the same patient cohort, the authors provided a controlled test of whether the peritumoral region adds predictive value. The significant performance gap between the two models suggests that the answer is yes, and that the tissue immediately around the tumor encodes information about treatment response that the tumor core does not. This aligns with a broader trend in cancer imaging research toward peritumoral and habitat-based analysis, in which algorithms map heterogeneity across multiple tissue zones rather than collapsing the tumor into a single region.

The authors are careful about the limits of their work, and appropriately so. An AUC of 0.71, while statistically superior to the tumor-only model, is not yet sufficient for high-stakes clinical decision-making on its own. The study was retrospective, drawing on public databases including The Cancer Imaging Archive, and the models were trained and validated within that framework rather than tested prospectively on new patients. The researchers themselves note that further prospective validation is required before the approach can be implemented clinically, and the identified radiomic patterns should be regarded as hypotheses to be confirmed in independent, multi-center cohorts rather than as ready-made clinical tools. Radiomics models also face well-known reproducibility challenges, since feature values can be sensitive to scanner type, imaging protocols and segmentation methods, all of which must be standardized before widespread deployment.

Even with those caveats, the study adds a meaningful piece to a rapidly evolving puzzle. If validated, a radiomics-based prediction tool could be applied to the MRI scans that patients with locally advanced breast cancer already undergo before treatment, adding no additional procedures, biopsies or costs. It could help identify patients unlikely to benefit from standard neoadjuvant chemotherapy early, opening the door to escalated or alternative strategies, while sparing predicted responders from unnecessary treatment modifications. It could also refine treatment stratification within molecularly distinct subgroups such as triple-negative and HER2-positive disease, where the stakes of getting the initial approach right are highest. The deeper promise lies in the conceptual shift the study embodies: treating the tumor and its microenvironment as a single, imageable system whose heterogeneity can be quantified, modeled and translated into decisions. As machine learning methods mature and imaging datasets grow, the invisible landscape around a tumor may prove to be just as informative as the tumor itself, and studies like this one are mapping the terrain.

Subject of Research: MRI radiomics for predicting pathological complete response to neoadjuvant chemotherapy in stage II/III breast cancer

Article Title: Enhancing pathological complete response prediction in stage II/III breast cancer: the role of radiomics signatures of MRI and its association with tumor microenvironment heterogeneity

Article References: Zheng, C., Sheng, J., Yang, L., Wang, M., Luan, X., Zhang, C., Liu, L., Jin, L., Zhao, Q., Fu, S., & Wang, B. (2026). Enhancing pathological complete response prediction in stage II/III breast cancer: the role of radiomics signatures of MRI and its association with tumor microenvironment heterogeneity. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02842-x

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02842-x

Keywords: breast cancer, radiomics, MRI, machine learning, pathological complete response, neoadjuvant chemotherapy, tumor microenvironment, tumor heterogeneity, peritumoral region, triple-negative breast cancer, HER2-positive, predictive medicine