Breast cancer remains one of the most prevalent and deadly cancers among women worldwide, with early detection being critical to improving patient outcomes. Mammograms, the frontline imaging tool for breast cancer screening, have traditionally been interpreted manually or with rudimentary computer assistance, primarily focusing on visible abnormalities. However, emerging deep learning (DL) techniques promise to decode subtler imaging biomarkers—patterns and features within the breast tissue that may signal an increased risk of cancer years before clinical manifestation.

The current study centers on Mirai, a cutting-edge deep learning model trained exclusively on mammographic images to predict breast cancer risk. Mirai harnesses complex patterns that escape conventional radiological assessment, extracting probabilistic risk estimations from raw imaging data. While Mirai has already demonstrated impressive predictive power, the researchers posited that integrating genetic data could elevate this approach. Specifically, they explored the incorporation of polygenic risk scores (PRS), which aggregate the effect of thousands of common genetic variants associated with breast cancer susceptibility.

Polygenic risk scores have emerged as a powerful genomic tool that quantify inherited cancer predisposition across a continuum of risk rather than relying on rare high-penetrance mutations alone. PRS capture immense genetic complexity and have been validated for breast cancer risk stratification in diverse populations. However, PRS alone do not provide spatial or temporal resolution about tissue changes, which imaging biomarkers uniquely offer. The fusion of these two complementary risk layers—image-based phenotypic risk and genotype-based inherited risk—conceptually promises a paradigm shift towards holistic, multi-modal risk prediction.

In this rigorous evaluation, the research team utilized a large, well-characterized cohort with comprehensive mammographic imaging and genotype data. They applied Mirai to mammograms to generate personalized risk estimates, then integrated these with independently derived polygenic risk scores calculated from participants’ genome-wide variant data. The integration was designed to assess additive or synergistic improvements in prediction accuracy, calibrated risk stratification, and clinical applicability.

The results, published in the British Journal of Cancer, confirm that while Mirai’s imaging-only predictions are robust, the addition of polygenic risk scores enhances discriminatory ability modestly but consistently. This finding is critical because even incremental gains in early risk prediction translate to significant clinical impact in terms of screening intervals, preventive interventions, and resource allocation. The combined model showed superior stratification of individuals into meaningful risk categories compared to either modality alone.

A key technical insight underpinning this success lies in the complementary nature of data sources. Mammographic images encode phenotypic manifestations of risk that can result from hormonal, environmental, or aging-related influences, while polygenic risk scores reflect inherited susceptibility embedded within the genome. By applying advanced probabilistic modeling techniques, the researchers effectively merged heterogeneous data to produce a unified risk estimate with enhanced predictive confidence.

Moreover, the study delved into how the integrated model performs across subpopulations, including different age groups, breast density categories, and ancestral backgrounds. Encouragingly, the combined approach maintained its performance robustness, suggesting broad clinical utility. This addresses a persistent challenge in breast cancer risk prediction—ensuring equitable accuracy across population strata commonly underrepresented in genomic and imaging datasets.

The implications of integrating AI-driven imaging biomarkers with polygenic risk extend well beyond risk estimation. The framework sets a precedent for multi-modal precision medicine where imaging, genomics, and potentially other data types like blood biomarkers or lifestyle factors can be cohesively analyzed. This could revolutionize how screening programs are personalized, enabling dynamic adjustment of screening frequency and modality based on evolving composite risk profiles.

Nevertheless, the authors acknowledge limitations and important areas for future investigation. The study cohort, while large, primarily represented populations of European ancestry, necessitating validation in more diverse ethnic groups given variability in genetic architecture. Additionally, the incremental performance boost, though statistically significant, underscores the need for further refinement in fusion algorithms and exploration of additional biomarkers to maximize predictive gains.

The integration of deep learning mammographic models with polygenic risk scores exemplifies a transformative trend in oncology—leveraging the power of AI and genomics to move beyond binary disease classification towards nuanced, individualized risk landscapes. It heralds a future where women can receive personalized breast cancer screening schedules tailored not only to imaging findings but also to their unique genetic risk, enabling earlier interventions that could substantially reduce morbidity and mortality.

This study lays a crucial foundation for clinical translation, emphasizing the potential for integrated multi-modal risk prediction tools to become standard components of breast cancer prevention strategies. As computational and genomic technologies continue to advance, we can expect progressively refined models that incorporate even deeper layers of biological complexity, from tumor microenvironment imaging to epigenetic modifications.

Importantly, this research underscores the need for collaborative efforts bridging radiology, genetics, data science, and clinical oncology to develop, validate, and implement these sophisticated predictive models in real-world healthcare settings. Ensuring interpretability, ease of integration into clinical workflows, and equitable access will be paramount challenges as these tools transition from research to practice.

Ultimately, the combination of imaging biomarkers and polygenic risk scores represents a monumental leap towards personalized oncology. It exemplifies how contemporary medicine harnesses massive data, sophisticated algorithms, and biological insights to tackle one of the most pressing health issues faced by women worldwide. The promise of AI-augmented genomic medicine in breast cancer risk prediction is not just to improve statistics but to transform lives through earlier, wiser, and more individualized care.

As breast cancer prevention enters this new era, the synergy between human biology and machine intelligence will redefine what is possible in early detection and precision intervention. This landmark study by Azam et al. thus serves as both a scientific milestone and a beacon guiding future exploration at the nexus of medical imaging and genomics, sparking renewed hope in the global fight against breast cancer.

Subject of Research: Breast cancer risk prediction using deep learning mammographic models combined with polygenic risk scores.

Article Title: Performance of an image-only deep learning breast cancer risk model with the addition of a polygenic risk score.

Article References:
Azam, S., Lamb, L.R., Eliassen, A.H. et al. Performance of an image-only deep learning breast cancer risk model with the addition of a polygenic risk score. Br J Cancer (2026). https://doi.org/10.1038/s41416-026-03415-z

Image Credits: AI Generated

DOI: 06 April 2026

The clinical challenge addressed by this research lies in the unpredictable nature of high-risk breast lesions. While some of these lesions remain benign, others upgrade to malignancy upon surgical excision, creating a critical need for improved non-invasive diagnostic tools to stratify patient risk effectively. Traditional imaging and biopsy methods have exhibited limitations in accurately forecasting these pathological upgrades, leading to unnecessary surgeries or delayed treatment. The innovative use of radiomics — extracting vast quantitative data from medical images — provides a promising avenue to overcome these constraints.

This study leveraged dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), a sophisticated imaging modality that highlights blood flow differences within breast tissue, to derive both clinicoradiological characteristics and complex radiomic features. By delineating both intratumoral and peritumoral regions of interest (ROIs), the team hypothesized that combining data from tumor cores and their immediate microenvironment could capture crucial biological interactions indicative of malignancy risk.

Data from 174 patients with biopsy-confirmed high-risk breast lesions were retrospectively analyzed. These patients underwent preoperative MRI scans between 2019 and 2024 at Shenzhen People’s Hospital. To ensure robust model development and validation, the dataset was split into training and test sets at a 7:3 ratio. The high granularity enabled the researchers to build multiple radiomic models focusing on various spatial regions — the lesion itself and peritumoral areas extended outward by 3 mm, 5 mm, and 7 mm respectively.

Particularly striking was the performance of the peritumoral 3 mm radiomics model, which outperformed broader surrounding regions. This suggests that the immediate peritumoral microenvironment harbors critical imaging features that correlate with risk of morphological upgrade. These findings highlight the importance of not only looking within the tumor boundaries but also closely analyzing its proximal tissue milieu, which may reflect early infiltrative or reactive processes preceding malignant transformation.

Beyond isolated radiomic signatures, the study further integrated clinical and conventional imaging features, combining them with both intratumoral and peritumoral radiomics to construct a comprehensive predictive model. This dual-region combined model demonstrated exceptional diagnostic performance, achieving an AUC (area under the curve) of 0.883 in the training cohort and 0.851 in the independent test set. These metrics significantly surpassed the predictive power of models based solely on clinical data or individual radiomic features.

Diagnostic sensitivity, specificity, and accuracy of the combined model were also impressive. In the training group, these metrics were 79.4%, 82.7%, and 81.8% respectively, while the test cohort exhibited 72.7% sensitivity, 85.7% specificity, and 83.0% accuracy. Such balanced performance underscores the model’s potential utility in real-world clinical settings where minimizing both false positives and negatives is paramount to patient outcomes and healthcare resource optimization.

The researchers employed rigorous univariate and multivariate logistic regression analyses to identify independent risk factors for pathological upgrade. These statistical approaches ensured the integration of only the most relevant radiomic and clinical features into the final model, mitigating overfitting and enhancing generalizability. The result is a nuanced risk stratification tool anchored in biologically meaningful data representation.

This research also culminated in the design of a clinically applicable nomogram — a graphical calculation tool that synthesizes multiple predictive factors into an individualized risk score. Such a nomogram can provide oncologists and radiologists with an intuitive interface to estimate upgrade probabilities, guiding personalized treatment decisions, such as whether to proceed with surgical excision or adopt a watchful waiting strategy.

Importantly, the study’s retrospective multicenter framework and relatively large sample bolster confidence in the findings, though prospective validation across diverse populations remains essential before widespread clinical adoption. The methodology, based on automated ROI delineation and multi-scale radiomic feature extraction, lays a replicable foundation for future investigations into other cancer types and lesion-risk assessments.

From a technological standpoint, the dual-region radiomic approach breaks new ground by recognizing the peritumoral environment as a critical player in oncogenesis and tumor progression. This paradigm shift broadens the imaging biomarker landscape and reflects trends in tumor microenvironment research, which increasingly reveal how surrounding stromal and immune components influence cancer behavior.

Given the rapid evolution of artificial intelligence and machine learning algorithms in medical imaging, this study exemplifies how advanced computational analytics can enable precision oncology. By harnessing subtle imaging textures, shape descriptors, and signal intensity variations imperceptible to the human eye, radiomics enhances diagnostic accuracy and unlocks new insights into tumor biology.

The clinical implications are profound. Accurate preoperative risk assessment helps avoid overtreatment in patients with benign high-risk lesions and conversely ensures timely intervention for those on the verge of malignant transformation. Moreover, this approach can reduce patient anxiety, limit unnecessary invasive procedures, and optimize healthcare resource allocation.

In a broader context, such advances contribute to the shifting landscape from “one-size-fits-all” cancer care to individualized management protocols based on precise phenotypic information. Incorporating quantitative radiomic signatures with clinical parameters exemplifies the future of multi-omic integration, potentially paving the way for more personalized, data-driven diagnostic and therapeutic pathways.

As the cancer research community continues to unravel the complexity of tumor heterogeneity and its clinical ramifications, studies like this underscore