Breast cancer is the most common cancer in women globally. Mammography has been the cornerstone of breast cancer screening since the 1960s, but its ability to accurately differentiate benign from malignant lesions remains limited -- especially in women with dense breast tissue.
Contrast-enhanced mammography (CEM) improves on standard mammography by using an intravenous iodinated contrast agent that highlights areas of abnormal blood vessel growth (angiogenesis) associated with malignant tumors. CEM has reported sensitivity of 90-100%, substantially higher than standard mammography or ultrasound alone, but its specificity of only 58-70% means a large proportion of flagged lesions are ultimately benign.
Radiomics -- the extraction and analysis of quantitative image features -- offers a potential solution to improve specificity. By mathematically describing the texture, intensity distribution, and spatial patterns within a tumor region, radiomics can capture information that may not be visible to the human eye, providing a more objective and reproducible characterization of the lesion.
This prospective single-institution study enrolled 80 patients with known breast lesions at a cancer institute in Naples, Italy, between October 2017 and April 2018. The goal was to test whether textural features extracted from CEM images could accurately classify lesions as benign or malignant using both univariate and multivariate AI-based methods.
A unique feature of this study was the use of three mammographic views: the standard cranio-caudal (CC) view and two time-points of the mediolateral oblique (MLO) view -- one acquired 4 minutes after contrast injection (early-MLO) and one at 8 minutes (late-MLO). This temporal dimension captures how contrast enhancement evolves over time, which may reflect differences in tumor vascularity between benign and malignant lesions.
In total, 104 lesion samples from 80 patients were analyzed -- 65 malignant and 39 benign -- using 48 standardized textural features extracted from manually segmented lesion regions on the subtracted CEM images.
Eligible patients had histologically confirmed unifocal or multifocal breast lesions and underwent dual-energy CEM on a Hologic Selenia Dimensions system. Contrast medium was injected intravenously at 1.5 mL/kg body weight, followed by CC-view imaging starting approximately two minutes later, with MLO-view images acquired at 4 and 8 minutes post-injection.
Each CEM acquisition produced a low-energy image equivalent to a standard digital mammogram, plus a recombined subtracted image that highlights contrast uptake. All textural feature extraction was performed on the subtracted images, where contrast enhancement is most clearly visible and lesion boundaries are better delineated.
Patients were excluded if they had contraindications to contrast agents, breast implants, inability to remain still, or had received neoadjuvant therapy prior to imaging. The exclusion criteria ensured that the radiomics analysis reflected untreated lesion characteristics, which is important for establishing baseline diagnostic performance.
Two expert radiologists with 15 and 22 years of experience independently segmented the region of interest (ROI) around each lesion on the subtracted images. Segmentation agreement was assessed using the intra-class correlation coefficient (ICC), which reached an excellent median value of 0.92, indicating high reproducibility of the texture measurements.
Using a standardized MATLAB toolbox based on the Image Biomarker Standardization Initiative (IBSI), 48 textural features were extracted per lesion per view. These features spanned four mathematical frameworks: first-order gray-level statistics (e.g., mean, variance, interquartile range), Gray-Level Co-Occurrence Matrix (GLCM) features capturing pixel-pair intensity relationships, Gray-Level Run-Length Matrix (GLRLM) features measuring intensity runs across directions, and Neighborhood Gray-Tone Difference Matrix (NGTDM) features quantifying local intensity variation including coarseness, busyness, and strength.
Five machine learning classifiers were tested: linear discriminant analysis (LDA), decision tree (DT), k-nearest neighbors (KNN), artificial neural network (NNET), and support vector machine (SVM). Feature selection was performed using LASSO regression, and class imbalance between benign and malignant cases was addressed using two synthetic oversampling methods: ADASYN and SASYNO.
Univariate analysis tested each texture feature independently using the Wilcoxon-Mann-Whitney test and ROC curve analysis. The best-performing individual features from the CC view were STRENGTH (AUC 0.74), COARSNESS (AUC 0.73), and VARIANCE and MAD (AUC 0.72 each) -- all from the Neighborhood Gray-Tone Difference Matrix or first-order statistics.
Features like STRENGTH and BUSYNESS quantify how locally variable the intensity patterns are across the lesion -- concepts related to tissue heterogeneity. Malignant tumors tend to show more irregular, heterogeneous enhancement patterns compared to the smoother, more uniform appearance of benign masses.
Despite statistical significance (all p-values below 0.001), the univariate AUC values of 0.70-0.74 across all three views represent only modest discriminative ability. These results confirm that no single textural feature is sufficient to reliably classify breast lesions, motivating the use of multivariate combination approaches.
For the CC view alone, the best multivariate model was a decision tree trained on three LASSO-selected features (MAD, VARIANCE, and LRLGE) with ADASYN balancing, achieving accuracy of 0.75 and AUC of 0.82. This minimal three-feature set offers an interpretable, parsimonious model that could be practically implemented.
The early-MLO view performed best with an artificial neural network trained on ten features including MEAN, MAD, VARIANCE, CORRELATION, and BUSYNESS, achieving accuracy of 0.77 and AUC of 0.83. The late-MLO view alone reached a maximum AUC of 0.82 using eleven features, suggesting that temporal enhancement dynamics add incremental but limited discriminative value compared to the initial enhancement captured in CC and early-MLO.
The highest performance was achieved by combining all 144 features (48 per view across three views) with an SVM classifier using ADASYN balancing and leave-one-out cross-validation: accuracy 0.87, AUC 0.90, sensitivity 0.86, and specificity 0.87. This substantial improvement over single-view analysis confirms that multi-view, multi-temporal information is significantly more informative than any single acquisition alone.
The dataset had 65 malignant and 39 benign lesions -- a common imbalance in oncology datasets reflecting the higher prevalence of malignancy in referred patients. Without addressing this imbalance, classifiers tend to predict the majority class (malignant) disproportionately, producing artificially high overall accuracy but poor specificity for benign cases.
ADASYN (Adaptive Synthetic Sampling) generates synthetic training examples in the minority class (benign), weighted by how difficult those cases are to classify -- focusing generation on the most challenging boundary cases. SASYNO (Self-Adaptive Synthetic Over-sampling) similarly generates synthetic minority examples but using adaptive neighborhood structures. Both approaches significantly improved model specificity compared to training on unbalanced data.
LASSO (Least Absolute Shrinkage and Selection Operator) feature selection penalizes model complexity by forcing less useful feature coefficients toward zero, retaining only the most informative predictors. This reduces overfitting and produces parsimonious models with 3 to 23 features depending on the view and lambda parameter -- substantially fewer than the full 48-feature set, and thus more generalizable to new patients.
This study's best result of AUC 0.90 using all three views compares favorably with prior CEM radiomics literature. Fanizzi et al. reported a fully automated CEM support system with sensitivity of 87.5% and specificity of 91.7%, while Losurdo et al. achieved 80% accuracy using an SVM on Haar-decomposed CEM features. The current study's results are competitive with these approaches while using a broader and more standardized feature set.
Marino et al. and La Forgia et al. demonstrated that CEM radiomics can predict molecular subtypes such as HER2 status, hormone receptor status, and tumor grade -- extending the role of quantitative imaging beyond simple benign/malignant classification. These studies suggest that textural features capture biologically meaningful information about tumor heterogeneity at a molecular level.
A key advantage of automated radiomics approaches is their potential immunity to confounding factors such as breast density and background parenchymal enhancement (BPE), which can limit radiologist interpretation. Because texture analysis operates on quantitative pixel-level data rather than subjective visual impressions, it may maintain diagnostic reliability in complex imaging scenarios where human readers struggle.
Currently, definitive diagnosis of breast malignancy requires biopsy -- an invasive, uncomfortable, and sometimes anxiety-inducing procedure. A radiomics model with sufficient specificity could identify patients with high-confidence benign lesions, reducing the biopsy rate while maintaining high cancer detection sensitivity.
The CEM imaging protocol used in this study -- CC view and two timed MLO acquisitions -- provides more temporal information than standard mammography and may approach the dynamic characterization offered by breast MRI, at lower cost and shorter scan time. Combining this functional information with quantitative texture analysis could create a practical, affordable alternative for dense-breast or high-risk screening programs.
Looking forward, integration of textural CEM features with histopathological data could improve grading and subtype classification -- moving beyond the binary benign/malignant distinction toward more nuanced characterization of invasiveness, hormone receptor status, and tumor biology, all of which directly influence treatment decisions.
This study demonstrated that multivariate AI analysis of CEM textural features -- especially when combining features across all three mammographic views -- can classify breast lesions as benign or malignant with accuracy of 0.87 and AUC of 0.90. This represents a meaningful improvement over both individual features and single-view models.
Key limitations include the small cohort of 80 patients and 104 lesions from a single institution, which limits statistical power and generalizability. Manual lesion segmentation introduces potential operator variability and is impractical at scale. The study also did not compare CEM radiomics performance directly against standard digital mammography, making it difficult to quantify the added value of contrast enhancement alone.
Future work will focus on expanding the cohort, implementing automatic lesion segmentation, including histopathological grading and molecular subtype classification in the analysis, and validating findings across multiple centers. These steps are essential to establish whether CEM-based radiomics can be translated into a clinically deployable, reproducible diagnostic tool.