Breast cancer is one of the leading causes of cancer-related death among women globally. Early detection through medical imaging significantly improves treatment outcomes, but existing modalities each carry limitations in sensitivity, specificity, or interpretive variability.
Contrast-enhanced mammography (CEM) is a newer technique that combines standard digital mammography with contrast imaging. After injecting an iodinated contrast agent, CEM captures two image types per view: a low-energy (LE) image resembling a standard mammogram, and a recombined (REC) image that highlights areas of high contrast uptake -- a marker of tumor angiogenesis. CEM has been shown to outperform standard mammography, especially in women with dense breast tissue.
Despite its promise, most AI tools for CEM are constrained by small datasets (typically a few hundred patients) and require manual identification of regions of interest before analysis. Few studies have built fully automated, large-scale detection systems that also account for variable clinical factors such as background parenchymal enhancement (BPE) -- the normal contrast uptake of healthy breast tissue that can obscure or mimic cancer signals.
This study from GE HealthCare and the Baheya Foundation in Egypt developed and evaluated a deep learning-based computer-aided detection (CEM-CAD) system for automated cancer detection and breast-level classification using CEM images. Unlike prior work, this system does not require manual region-of-interest extraction.
The study used a large, heterogeneous dataset of 1,673 patients and 7,443 image pairs collected from seven institutions across five countries. This multisite diversity provides a more realistic assessment of generalizability than single-institution studies.
Key contributions include: evaluating the impact of using LE images alongside REC images (versus REC alone), comparing model performance to published radiologist benchmarks from clinical review literature, and rigorously analyzing performance across different BPE levels -- a clinically significant factor previously understudied in CEM-CAD work.
The primary dataset came from the Baheya Foundation for Early Detection and Treatment of Breast Cancer in Giza, Egypt -- 1,087 patients with CEM images acquired from December 2019 to December 2022. All lesions were biopsy-proven, and annotations included lesion location boxes, breast density, BPE level, BI-RADS category, and pathology outcome.
A second dataset of 586 patients was collected from six additional institutions in China, the United States, Spain, and the United Kingdom, using multiple GE Healthcare imaging systems. This ensured that the model learned from images produced under diverse acquisition conditions and patient demographics.
The test set comprised 150 cases drawn exclusively from the Baheya dataset, stratified by pathology outcome (75 malignant, 50 biopsy-proven benign, 25 normal/non-biopsied). This stratification ensured a clinically representative test population. All non-biopsied cases were assumed benign under BI-RADS 1-2 classification, a standard assumption in screening-oriented studies.
The detection model was built on YOLOv5 (You Only Look Once, version 5), a single-stage object detection architecture known for speed and accuracy. Unlike two-stage detectors that first propose candidate regions and then classify them, YOLO processes the full image in a single pass, predicting bounding boxes and class probabilities simultaneously.
The YOLOv5 framework has three main components: a CSPDarknet backbone for feature extraction, a path aggregation network (PAN) neck that fuses multi-scale feature maps to handle lesions of varying sizes, and a YOLO detection head that outputs bounding box coordinates and class scores. The model was trained to predict three lesion categories: malignant, biopsy-proven benign, and non-biopsied findings -- though only the malignant class was reported in this evaluation.
To improve localization accuracy, a distance intersection over union (DIoU) loss function was used instead of standard IoU. DIoU adds a penalty term based on the distance between the centers of the predicted and ground-truth boxes, ensuring more stable optimization even when bounding boxes don't overlap initially.
Five models of varying sizes (YOLOv5s and YOLOv5m in two configurations, and YOLOv5l) were independently trained and then combined using model ensembling -- averaging predictions across all five -- to improve robustness and reduce variance in final outputs.
Performance was assessed at two levels. At the lesion level, the Free Receiver Operating Characteristic (FROC) curve plotted cancer detection sensitivity against the number of false positives per image. The primary summary metric was the area under the normalized FROC curve up to 0.3 false positives per image (AUFROC 0.3).
Lesion detection was validated using an Intersection over Union (IoU) threshold of 0.1 -- a relatively lenient criterion chosen to account for annotation variability across institutions, where the same multifocal tumor might be annotated as one large box or several smaller ones depending on the radiologist.
At the breast level, the model's overall suspicion score (based on the most suspicious detection in any view of a breast) was used to compute a standard AUROC curve comparing malignant versus non-malignant breasts. Performance was compared to reference radiologist results compiled from published meta-analyses of CEM diagnostic studies.
The ensemble model achieved a breast-level AUROC of 0.964 (95% CI: 0.946-0.979) when using both LE and REC images as inputs -- outperforming the prior best published CEM-CAD (AUROC 0.930) and matching the upper range of radiologist performance from clinical review literature.
At the lesion level, the model detected 90% of all cancers with only 0.128 false positive detections per image at an IoU threshold of 0.1. At the more lenient IoU threshold of 0.05 (which better accounts for annotation inconsistencies), this rate dropped to 0.085 false positives per image at the same 90% sensitivity -- a meaningful improvement over previously published systems.
Comparing to the median radiologist performance benchmark (specificity of 0.744) extracted from 24 clinical studies, the AI model achieved a sensitivity of 0.988 at that fixed specificity point. Individual radiologist sensitivities in those 24 studies ranged from 0.543 to 1.00, with a median of 0.952 -- placing the AI at or above the median human performance level.
Using both LE and REC images as model inputs substantially improved performance over using REC alone. Breast-level AUROC increased from 0.916 to 0.964, and the AUFROC for lesion detection improved from 0.733 to 0.853 -- a 16% relative gain. This confirms that the morphological information in low-energy images complements the functional contrast data in recombined images, mirroring how radiologists themselves use both image types in clinical review.
Background parenchymal enhancement (BPE) had a large and consistent impact on model performance. For patients with low BPE (minimal or mild), the AUROC was 0.986 and the lesion-level AUFROC was 0.891. For patients with high BPE (moderate or marked), performance dropped to AUROC 0.919 and AUFROC 0.712. The high false positive rate of 0.416 per image in the high-BPE subgroup illustrates how normal parenchymal enhancement can mimic malignant uptake.
This BPE impact is consistent with clinical literature showing that high BPE increases false-positive rates in CEM interpretation. The authors note this as a key area for future optimization, including targeted collection of more high-BPE training cases and development of BPE-aware algorithmic strategies.
The proposed model outperformed two comparable prior CEM detection systems. Compared to Jailin et al.'s previous detection model, the new system improved AUROC from 0.930 to 0.964, AUFROC from 0.632 to 0.853, and reduced false positives per image at 90% sensitivity from 0.410 to 0.128. Unlike Zheng et al.'s multicenter study (AUROC 0.947), this model additionally handled complex cases such as multiple lesions, bilateral lesions, non-mass lesions, and lesions visible in only one view.
A key limitation is that training and testing were performed primarily on data from a single institution (Baheya Foundation), which may introduce site-specific annotation and population biases. The model was trained by a single group of radiologists, increasing the risk of annotation-driven overfitting to particular interpretive styles.
Another limitation is the assumption that all non-biopsied cases are benign -- an approximation that could undercount rare false negatives in the training data. Additionally, the current model processes each image view independently, without leveraging cross-view consistency between cranio-caudal and mediolateral oblique projections, which radiologists routinely use to improve confidence in lesion identification.
The model's performance -- high sensitivity with a low false positive rate per image -- positions it as a practical tool for assisting radiologists in CEM interpretation. At 90% sensitivity with 0.128 false positives per image, the system would reliably flag most cancers while generating a manageable volume of false alarms.
The visual heatmap outputs provide spatial context for each detection, highlighting probable cancer regions in red and probable benign areas in blue. This type of output enables radiologists to quickly scan AI-flagged areas and either confirm or override detections, supporting an interactive human-AI workflow rather than requiring full automation.
One promising near-term application is integration with full-field digital mammography (FFDM) CAD systems, which have access to much larger training datasets. Combining an FFDM CAD on the low-energy images with this CEM-specific model on the recombined images could further improve discrimination, particularly for calcification-associated malignancies that are visible on LE but not REC images.
This study demonstrated that a YOLOv5-based AI model trained on a large, diverse CEM dataset can achieve radiologist-competitive performance for automated breast cancer detection and classification. The combination of LE and REC image inputs, model ensembling, and test-time augmentation produced the best overall results.
The investigation of BPE as a performance modifier is a particularly important finding, as it identifies a specific patient subgroup where current AI tools underperform and where additional development effort should be focused. Collecting more high-BPE training cases is a concrete next step.
Future directions include multi-institution expansion with standardized annotation protocols, multi-view model architectures that leverage left-right and CC/MLO view consistency, and prospective clinical validation to measure the actual impact of CEM-CAD on radiologist workflow efficiency and diagnostic accuracy in real-world settings.