Standard digital mammography reduces breast cancer deaths by approximately 40%, but it has a well-documented limitation: in women with dense breasts -- approximately 43% of women aged 40 to 85 -- overlapping fibroglandular tissue can hide small early-stage cancers. Contrast-enhanced mammography (CEM) was developed to overcome this. CEM injects an iodinated contrast agent intravenously, then uses a dual-energy imaging technique to create a subtracted image that suppresses normal breast tissue while highlighting areas of contrast uptake -- revealing lesions based on their blood supply rather than just density. Studies consistently find CEM's sensitivity comparable to MRI but at significantly lower cost and with higher patient throughput.
Despite these advantages, AI development for breast imaging has focused almost entirely on conventional digital mammography datasets. This means the AI models being developed, validated, and deployed are designed for a less sensitive modality, while the more capable technology remains largely without algorithmic support. The mismatch exists because CEM datasets are rare and small compared to digital mammography collections, creating the same data scarcity challenge that affects AI development across specialized medical imaging fields.
This review, covering 14 studies from 2018 to 2023 selected through a systematic PRISMA-guided search, examines the state of AI applied specifically to CEM images. It analyzes the full pipeline from image acquisition through preprocessing, segmentation, feature extraction, and classification, identifying both current achievements and the standardization gaps that must be addressed before AI for CEM can achieve clinical-scale deployment.
CEM produces two image types for each view: a low-energy image similar in appearance to standard digital mammography, and a recombined (subtracted) image that shows areas of contrast enhancement in a background-suppressed format. The dual-energy acquisition uses two distinct voltage ranges -- typically 26 to 34 kilovolts for the low-energy exposure and 45 to 49 kilovolts for the high-energy exposure -- with vendor-specific anode and filter materials. Among the 14 reviewed studies, 9 used GE Healthcare systems, 1 used Hologic, and 3 used data from both vendors.
A critical issue identified in the review is that no standardized implementation protocol has been established across the four major CEM vendors. Each vendor uses different tube voltage ranges, anode materials, filter thicknesses, and image reconstruction algorithms to create the subtracted recombined image. These differences can introduce systematic visual variation that is not biologically meaningful -- a form of technical noise that can bias AI models trained or tested across multiple vendors or institutions. This inter-vendor variability is a major source of what the review calls inter-operator bias in multicenter studies.
The review strongly recommends that publications disclose vendor, model, and acquisition protocol details, noting that several reviewed studies failed to provide this basic information. Without vendor information, it is impossible to evaluate whether performance differences between studies reflect genuine algorithmic advances or merely different imaging system characteristics. Studies using multivendor data should explore harmonization strategies -- techniques for standardizing images across different acquisition systems before training or testing AI models -- as a prerequisite to reliable multicenter AI research.
CEM images require several preprocessing steps before AI analysis. Background removal eliminates artifacts and non-breast regions that could mislead models. Methods range from simple thresholding (effective for large, well-defined artifacts) to clustering algorithms that group similar pixels (better for smaller or grouped artifacts) to deep learning models trained specifically to identify and remove artifacts. Two reviewed studies used Otsu's thresholding method, one in combination with Contrast Limited Adaptive Histogram Equalization (CLAHE), which improves local contrast while preventing over-enhancement in already-bright regions.
Resampling is required when images from different centers or machines have different pixel resolutions, as even slight resolution differences can introduce systematic feature differences that confuse trained models. Normalization addresses variation in brightness, contrast, and color balance across different acquisition protocols and contrast injection dosages. One study trained on data from one institution and tested on two others, using normalization to ensure dataset consistency across sources. Min-max normalization was used in another study after pixel value adjustment to improve contrast.
Data augmentation is particularly important for CEM because datasets are small and class-imbalanced. Techniques applied in reviewed studies include horizontal flipping, rotation, scaling, size reduction, and geometric transformations of breast shape. One study used realistic transformations of breast geometry specifically designed for mammographic data rather than generic image augmentation. GAN-based augmentation was also attempted in one study but the generated images did not satisfy expert radiologists, so traditional augmentation was used instead -- an important finding suggesting GAN approaches for CEM may require further development.
Two fundamentally different approaches to feature extraction emerged across the 14 reviewed studies. Handcrafted radiomics involves expert-specified mathematical features extracted from lesion regions that have been manually segmented: shape features (lesion dimensions and boundary characteristics), texture features (heterogeneity of pixel intensity), and kinetic features (changes in enhancement over time). Tools such as the PyRadiomics package and the IBSI-standardized Texture Toolbox in MATLAB were used. After extraction, features undergo redundancy elimination using methods such as LASSO regression, Spearman correlation analysis, and interobserver agreement testing.
Transfer learning uses pretrained deep neural networks -- ResNet, Inception V3, CSP DarkNet, VGG16, Xception, RetinaNet -- to automatically extract features from CEM images without requiring manual segmentation of lesions first. Six out of seven CNN-based studies used transfer learning, with ResNet being the most common backbone choice. Transfer learning is particularly suited to CEM because of its small available datasets: rather than learning from scratch, the model begins with visual features already learned from millions of other images and adapts them to CEM data.
The choice between the two approaches carries practical implications. Handcrafted radiomics requires lesion segmentation as a prerequisite, meaning errors in segmentation propagate into the features. Transfer learning can operate on entire images or lesion patches, bypassing the segmentation requirement. One study combined both approaches -- using handcrafted radiomics features alongside deep learning features -- and achieved superior performance to either method alone. The review recommends incorporating quantitative feature analysis even in deep learning studies to improve model interpretability and comparability across institutions.
The 14 reviewed studies collectively cover a range of AI approaches for CEM malignancy classification, from classical machine learning to deep CNN architectures. The best-performing studies achieved strong results: logistic regression combined with LASSO on radiomics features reached AUC of 0.96 with 94.6% accuracy and 97% sensitivity; a YOLOv5-based detection system achieved 90% accuracy and AUC of 0.964 on over 7,400 images; and an artificial neural network using radiomics features reached AUC of 0.97 with 95% sensitivity. Seven studies used machine learning for classification and seven used CNNs, with performance roughly comparable between the two approaches when properly applied.
Half the studies used machine learning techniques (logistic regression, SVM, random forest, linear discriminant analysis) applied to extracted radiomics features. These methods offer interpretability -- the clinician can examine which features drove each prediction -- but may not capture the most complex visual patterns. The other half used CNNs, which learn features automatically from image data and can capture subtle patterns invisible to manual feature engineering. Among CNN studies, 6 of 7 used transfer learning with pretrained networks, reflecting the practical reality that CEM datasets are too small to train deep networks from scratch.
The review identifies important reporting gaps across studies. Many CNN papers failed to specify hyperparameter choices such as learning rate schedule, optimization algorithm, batch size, dropout rates, and regularization methods. Several did not explain why particular objective functions were chosen. Without this information, it is impossible to reproduce or validate results independently, limiting the scientific value of the work. The CLAIM (Checklist for Artificial Intelligence in Medical Imaging) framework is cited as the appropriate standard for ensuring transparent and reproducible AI reporting in radiology.
Several studies demonstrated that incorporating patient clinical information alongside image features substantially improved AI performance. In one study, adding demographic and clinical data to image-based features produced a notably higher AUC than using any single image type alone. In another, combining clinical features with radiomics features from CEM images produced significant increases in both accuracy and sensitivity compared to either source alone. These findings suggest that CEM AI systems should be designed as multimodal systems from the outset, not as pure image analysis tools.
The most impactful future direction identified by the review is multimodal data fusion: combining CEM images with ultrasound, MRI, or other breast imaging modalities to exploit complementary information. CEM provides functional information about tumor vascularity; ultrasound characterizes tissue texture and mass consistency; MRI offers the highest soft tissue contrast and three-dimensional anatomical detail. A model that learns from all three could overcome the individual limitations of each while requiring fewer independent investigations per patient.
Handling class imbalance is identified as a persistent challenge across CEM studies. Recommended strategies include over-sampling techniques like SMOTE (Synthetic Minority Oversampling Technique) and ADASYN, which generate synthetic minority class examples to balance training data. Under-sampling is generally not recommended for CEM specifically because available datasets are already small, meaning removing majority class samples exacerbates the data scarcity problem. Cost-sensitive learning -- assigning higher penalties to errors on the rarer class -- provides an alternative that avoids modifying the dataset size.
AI for CEM has demonstrated genuine promise across all reviewed dimensions: image quality improvement, lesion detection, segmentation, radiomics feature extraction, and malignancy classification. The clinical case for CEM-AI is compelling -- improving on a modality that is already more sensitive than standard mammography, at lower cost than MRI, and with particular benefit for the 43% of women with dense breasts who are most poorly served by existing screening. Successful AI integration could reduce unnecessary biopsy procedures, improve early detection rates, and support consistent radiologist interpretation.
However, several barriers must be addressed before large-scale clinical adoption. Labeled training data for CEM remains scarce and few datasets are publicly available. Variability across imaging systems and acquisition protocols has not been adequately addressed through harmonization. Reporting standards for AI studies have not been consistently followed, limiting the ability to compare, reproduce, and build on existing work. And the studies reviewed are predominantly retrospective with limited external validation -- prospective multicenter studies confirming clinical value have not yet been completed.
The review concludes that AI in CEM represents a genuine frontier but one that remains in early development relative to conventional mammography AI. Addressing data scarcity through shared datasets, standardizing protocols across vendors, enforcing reporting requirements aligned with the CLAIM checklist, and designing prospective validation studies are the concrete steps needed. The synergistic potential between radiologists and AI is substantial -- but realizing it requires the collaborative effort of researchers, clinicians, and technology developers working toward shared standards.