A groundbreaking study published in BMC Cancer introduces a revolutionary multi-modal radiomics model designed to predict pathological complete response (pCR) to neoadjuvant treatment (NAT) in breast cancer patients. This pioneering approach integrates four distinct imaging modalities—ultrasound (US), mammography (MM), computed tomography (CT), and magnetic resonance imaging (MRI)—to significantly enhance the predictive accuracy of treatment outcomes. As neoadjuvant treatments become more prevalent in breast cancer management, accurately identifying patients likely to achieve pCR is paramount for optimizing therapeutic strategies and improving survival rates.
Radiomics, the practice of extracting high-dimensional quantitative features from medical images, has already proven its potential in oncology by advancing personalized medicine. However, prior radiomics models in breast cancer typically leveraged only a single imaging source. The innovative aspect of this study lies in combining the radiomic data derived from multiple imaging technologies, hypothesizing that a synchronized, multi-modal analysis would offer superior clinical insights. Integrating these diverse imaging datasets allows for a multifaceted evaluation of tumor heterogeneity and biological characteristics, which are often invisible to the naked eye or single modality assessments.
The research team conducted a retrospective analysis of 89 breast cancer patients who underwent surgery following NAT between January 2019 and July 2023. The patient cohort was characterized by a pCR rate of 31.5%, which aligns with typical response rates reported in similar clinical settings. By systematically extracting radiomic features from volumes of interest across US, MM, CT, and MRI scans, the study harnessed complex image texture, shape, and intensity data reflective of tumor microenvironment dynamics and structural changes induced by therapy.
A key methodological element was the application of the least absolute shrinkage and selection operator (LASSO), a regularization technique instrumental in selecting the most robust radiomic features while mitigating overfitting risks. This step ensured that the resulting radiomic signatures for each imaging modality were both predictive and generalizable. Subsequent statistical modeling combined these signatures into a comprehensive multi-modal radiomics framework, which was further enriched by incorporating independent clinical risk factors, namely progesterone receptor (PR) status, human epidermal growth factor receptor 2 (HER2) status, and clinical tumor (T) stage.
Notably, the study reported the area under the receiver operating characteristic curve (AUC) as the primary metric for model performance. Individual imaging modalities demonstrated moderate predictive power, with CT radiomics yielding the highest single-modality AUC of 0.814, followed closely by MRI at 0.787. Mammography and ultrasound lagged slightly behind, with AUCs of 0.762 and 0.702, respectively. These results underscore the variability inherent in each imaging technique’s capacity to capture therapy-induced tumor changes.
The real breakthrough emerged when the four radiomic signatures were amalgamated into a unified multi-modal radiomics model, achieving an impressive AUC of 0.904 and a Brier score of 0.111, indicating excellent calibration and predictive accuracy. Crucially, the addition of clinical risk factors propelled performance even further—the combined model attained an outstanding AUC of 0.943 alongside a Brier score of 0.082. This synergistic integration underscores the value of combining quantitative imaging biomarkers with established pathological and clinical indicators.
To translate these advancements into clinical utility, the investigators developed a nomogram visualizing the combined model. Nomograms serve as intuitive, user-friendly tools that enable clinicians to estimate the probability of treatment response on an individual basis, thus facilitating personalized therapeutic decisions. The availability of such a tool promises to bridge the gap between sophisticated computational models and everyday clinical practice.
The implications of this study are profound and multifold. Firstly, it challenges the prevailing paradigm of relying solely on single-modality imaging in radiomics research, providing compelling evidence for a multi-modal approach. By pooling diverse imaging features, the resultant model captures complementary tumor characteristics, such as metabolic activity, vascularization, and tissue density variations, all of which are essential to comprehensively understanding the tumor’s response to NAT.
Moreover, the inclusion of clinical variables alongside radiomic data highlights a paradigm shift towards fully integrated biomarker models. This holistic approach acknowledges that while imaging can reveal structural and functional insights, molecular markers like PR and HER2 status remain indispensable in defining tumor biology and treatment responsiveness. Such integration is essential to achieving the goal of precision oncology.
Technically, this study exemplifies the growing sophistication of machine learning techniques applied to medical imaging. The use of LASSO for feature selection and rigorous five-fold cross-validation for model validation reflects best practices in reducing bias and ensuring replicability. Reproducibility remains a crucial concern in radiomics, and this study’s methodological rigor provides confidence in the robustness of its findings.
Looking ahead, this study sets the stage for the development of broadly applicable, multi-modal radiomics platforms that can be deployed in clinical workflows. Future research may extend these findings by validating the model in larger, multicenter cohorts and exploring integration with genomic and proteomic data. Additionally, the model’s applicability to other cancer types treated with neoadjuvant therapies represents an exciting avenue for exploration.
The promising results garnered from CT and MRI modalities suggest a potential prioritization in clinical imaging protocols. However, the unique advantages of ultrasound and mammography, including accessibility and cost-efficiency, remain valuable, especially in diverse healthcare settings where advanced imaging may be limited.
Importantly, the adoption of such predictive models could transform therapeutic decision-making, enabling oncologists to tailor neoadjuvant regimens based on the likelihood of complete pathological response. This could minimize overtreatment and its associated toxicities, as well as identify patients who may benefit from alternative strategies early in the treatment course.
Furthermore, the development of such multi-modal radiomics models aligns with the overarching trend towards non-invasive biomarkers in oncology. Imaging-based predictive tools offer repeatable assessments without the risks and discomfort of biopsy procedures, fostering dynamic monitoring of treatment efficacy in real time.
In conclusion, this innovative study heralds a new era in breast cancer management, harnessing the full spectrum of imaging technology combined with clinical insights to precisely predict treatment outcomes. As the oncology community moves towards increasingly personalized approaches, multi-modal radiomics models such as this will undoubtedly become invaluable assets in the clinician’s armamentarium, ultimately improving patient prognosis and quality of life.
Subject of Research: Prediction of pathological complete response to neoadjuvant treatment in breast cancer using a multi-modal radiomics model.
Article Title: Multi-modal radiomics model based on four imaging modalities for predicting pathological complete response to neoadjuvant treatment in breast cancer.
Article References:
Liang, Y., Xu, H., Lin, J. et al. Multi-modal radiomics model based on four imaging modalities for predicting pathological complete response to neoadjuvant treatment in breast cancer. BMC Cancer 25, 985 (2025). https://doi.org/10.1186/s12885-025-14407-2
Image Credits: Scienmag.com
