# Contrast-Enhanced Mammography and Deep Learning-Derived Malignancy Scoring in Breast Cancer Molecular Subtype Assessment

## TL;DR

This study explores how contrast-enhanced mammography and AI can help identify breast cancer subtypes by analyzing tumor features and malignancy scores.

## Contribution

The study introduces the use of deep learning-derived malignancy scores in contrast-enhanced mammography for breast cancer molecular subtype assessment.

## Key findings

- Luminal tumors were more often spiculated with heterogeneous enhancement, while HER2-positive/triple-negative tumors were round with homogeneous enhancement.
- Deep learning scores showed higher median values for malignant lesions compared to benign ones, with a significant difference observed.
- AI scores varied across subtypes but differences were not statistically significant.

## Abstract

Background and Objectives: Contrast-enhanced mammography (CEM) provides both morphological and functional information and may reflect breast cancer biology similarly to Magnetic Resonance Imaging (MRI). Materials and Methods: This single-center retrospective study included 399 women with Breast Imaging Reporting and Data System (BI-RADS) category 0 screening mammograms who subsequently underwent CEM. A total of 76 malignant lesions (68 invasive cancers, 8 ductal carcinoma in situ (DCIS)) with complete imaging and pathology data were analyzed. Invasive cancers were classified into luminal A, luminal B, luminal B/Human Epidermal Growth Factor Receptor 2 (HER2)-positive, HER2-enriched, and triple-negative, and grouped as luminal (Group 1) versus HER2-positive/triple-negative (Group 2). Results: Luminal subtypes predominated (47 of 68, 69%), while 21 of 68 (31%) were HER2-positive or triple-negative. Most cancers appeared as masses with spiculated margins and heterogeneous enhancement. Significant differences were observed in mass shape (p = 0.03) and internal enhancement (p = 0.01). Luminal tumors were more often irregular and spiculated with heterogeneous enhancement, whereas the HER2-positive/triple-negative tumors more frequently appeared round with rim or homogeneous enhancement. Deep learning-derived malignancy scores (iCAD ProFound AI®) demonstrated good diagnostic performance (area under the curve (AUC) = 0.744, 95% confidence interval (CI) 0.654–0.821, p < 0.001). The median AI score was significantly higher in malignant compared with benign lesions (70% [interquartile range (IQR) 47–93] vs. 38% [IQR 25–61]; Mann–Whitney U test, p < 0.001). Among malignant lesions, iCAD scores varied across molecular subtypes, with higher median values observed in Group 1 versus Group 2 (87% vs. 55%), although the difference was not statistically significant (Mann–Whitney U test, p = 0.35). Conclusions: CEM features mirrored subtype-specific phenotypes previously described with MRI, supporting its role as a practical tool for enhanced tumor characterization. Although certain imaging and AI-derived parameters differed descriptively across subtypes, no statistically significant differences were observed. As deep-learning models continue to evolve, the integration of AI-enhanced CEM into clinical workflows holds strong potential to improve lesion characterization and risk stratification in personalized breast cancer diagnostics.