Every year, millions of women undergo breast imaging that reveals a finding of uncertain significance—a shadow on a mammogram, a mass on an ultrasound—and face the same anxious question: is this cancer or not? The standard answer often involves biopsy, an invasive procedure that most of the time, for the majority of patients, ultimately confirms a benign lesion. A new multicenter study published in BMC Medical Imaging suggests that artificial intelligence, working on two complementary imaging modalities, could help shrink that pool of unnecessary biopsies while catching the malignancies that matter.
The research, led by Dong-Man Ye, Si-Meng Song, and Peng Lu of the Cancer Hospital of Dalian University of Technology, together with colleagues at the Affiliated Hospital of Qingdao University, tackles one of the most persistent challenges in breast radiology: the subjective interpretation of conventional imaging. Radiologists reading digital breast tomosynthesis (DBT) and ultrasound (US) images must judge whether a lesion’s appearance—its margins, its internal texture, its shadowing behavior—points toward benign or malignant disease. That judgment varies between readers and even between readings by the same person. The study’s premise is that quantitative image analysis, known as radiomics, can extract objective information that the human eye either misses or cannot systematically weigh.
Radiomics works by converting medical images into hundreds or thousands of numerical descriptors. First-order statistics capture the distribution of pixel intensities within a lesion. Shape features quantify geometry—roundness, compactness, surface irregularity. Texture features describe spatial relationships between pixels at multiple scales, encoding patterns such as heterogeneity that pathologists recognize under the microscope as hallmarks of tumor architecture. In this study, the researchers also employed a more sophisticated variant called habitat radiomics, which partitions each lesion into distinct spatial subregions—habitats—based on local intensity characteristics, then computes features within and across those subregions. The idea mirrors tumor biology: cancers are not uniform masses but heterogeneous ecosystems, and their internal spatial organization carries diagnostic information.
The team built four kinds of models. A clinical model used the BI-RADS category assigned by radiologists under the American College of Radiology’s Breast Imaging-Reporting and Data System, the standard framework for rating suspicion on breast imaging. A radiomics model used handcrafted features extracted from regions of interest delineated by experienced radiologists on both DBT and US images. A habitat model used the subregion-based features. And a deep learning model extracted features automatically using a pretrained ResNet101, a convolutional neural network originally trained on millions of natural images and widely repurposed in medical AI. Feature selection employed the least absolute shrinkage and selection operator (LASSO), a statistical technique that shrinks irrelevant features toward zero, and the final classifiers were logistic regression models. Performance was measured by the area under the receiver operating characteristic curve (AUC), with the DeLong test used to compare models statistically.
The results are striking. In the training cohort, the clinical model achieved an AUC of 0.943, the habitat model 0.965, and the deep learning model 0.926—all significantly outperforming the conventional radiomics model, which reached only 0.826. No statistically significant differences separated the clinical, habitat, and deep learning models from one another. In the external test cohort drawn from the second center, the clinical model and the deep learning model again significantly outperformed the radiomics model, with AUCs of 0.943 and 0.874 respectively. The deep learning model’s robust performance on data from an institution it had never seen is particularly noteworthy, because external validation is where many AI models in medicine falter.
The multicenter design deserves emphasis. The study included 183 patients from the first center and 65 from the second, a total of 248 lesions. Training and testing on data from different hospitals guards against a well-known failure mode of machine learning in medicine: overfitting to scanner-specific artifacts, patient demographics, or radiologist habits at a single institution. The fact that the deep learning model maintained an AUC near 0.87 on unseen external data suggests the model learned features that generalize, not quirks of one imaging protocol.
Why combine DBT and ultrasound? The two modalities see different things. DBT, a three-dimensional evolution of digital mammography, reconstructs the breast in thin slices, reducing tissue overlap that obscures lesions on conventional two-dimensional mammograms. It excels at revealing architectural distortion, spiculated margins, and calcifications. Ultrasound, meanwhile, characterizes tissue stiffness and internal echo patterns in real time, distinguishing cysts from solid masses and revealing posterior acoustic features—shadowing or enhancement—that correlate with lesion composition. Neither modality alone captures the full picture. By extracting features from both and fusing them, the models effectively performed a virtual multimodal assessment, quantifying the same complementary information that expert radiologists integrate intuitively.
The clinical implications center on the biopsy decision. BI-RADS category 4 lesions—those with moderate suspicion—are the classic gray zone. Guidelines recommend biopsy for these findings, yet the majority prove benign on pathology. A quantitative tool that could reliably stratify BI-RADS 4 lesions into lower- and higher-risk subgroups could spare many women an invasive procedure, with its attendant anxiety, cost, pain, and small but real complication risks. The study’s authors frame their models as adjunctive tools—decision support for radiologists rather than replacements—designed to support lesion characterization and reduce unnecessary biopsies. Decision curve analysis, included in the study, evaluates the net clinical benefit of a model across a range of threshold probabilities, addressing exactly this question of when a model’s predictions translate into better decisions.
The study also highlights a subtle but important finding: the simple clinical model built on BI-RADS categories performed remarkably well, matching or exceeding the sophisticated image-based models in several comparisons. This is a recurring lesson in medical AI. The BI-RADS system already distills decades of accumulated radiological knowledge into a single ordinal score, and any new tool must beat it, not merely match it, to justify adoption. That the habitat and deep learning models matched but did not significantly exceed the clinical model suggests that the quantitative approaches capture information substantially overlapping with expert judgment—valuable for standardization and for reducing reader variability, but not yet a decisive leap forward.
Limitations remain, as the authors acknowledge. The study was retrospective, and retrospective designs are vulnerable to selection bias and to the availability of only lesions that already underwent biopsy, which can inflate apparent diagnostic performance. The sample size, particularly in the external cohort, was modest. The models were built on manually delineated regions of interest, adding a layer of human input that fully automated segmentation would remove. And the deep learning model relied on a network pretrained on natural images rather than on breast-specific data, leaving room for improvement with domain-adapted architectures. The authors call for prospective multicenter studies to validate the approach before clinical implementation.
Still, the trajectory is clear. Radiomics and deep learning are converging on the same goal: extracting more diagnostic information from the images that patients already undergo, at no additional cost, radiation, or invasiveness. The combination of DBT and US radiomics represents a pragmatic path toward precision breast imaging—one that works with existing scanners, existing workflows, and existing clinical questions. If future prospective studies confirm these findings, the anxious wait after an abnormal breast image could increasingly end not with a biopsy needle but with a confident, quantitative, and noninvasive answer.
Subject of Research: Radiomics analysis based on digital breast tomosynthesis and ultrasound for differentiating benign from malignant breast lesions
Article Title: Radiomics analysis based on DBT and US for differentiating benign from malignant breast lesions: a multicenter study
Article References: Ye, D.-M., Song, S.-M., Lu, P., Cui, C.-X., Gao, K.-R., & Yu, T. (2026). Radiomics analysis based on DBT and US for differentiating benign from malignant breast lesions: a multicenter study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02704-6
Image Credits: AI Generated
DOI: 10.1186/s12880-026-02704-6
Keywords: radiomics, digital breast tomosynthesis, ultrasound, breast lesions, deep learning, BI-RADS, machine learning, medical imaging, analysis, based, differentiating, benign
News Source: Ophelia Keating. (October 19, 2026). AI Model Combines DBT and Ultrasound Radiomics to Distinguish Benign from Malignant Breast Lesions. Scienmag.
