Standard digital mammography (DM) is the gold standard for breast cancer screening but has a well-known limitation: in women with dense breast tissue, fibroglandular tissue appears white on the image -- the same color as masses and tumors -- making cancers harder to see. Contrast-enhanced spectral mammography (CESM) was developed specifically to address this. Approved by the FDA in 2011, CESM uses an intravenous iodinated contrast agent that accumulates in blood vessels feeding rapidly growing tumors. This contrast uptake is then made visible through a specialized imaging technique.
CESM works by acquiring two exposures for each view: a low-energy image that resembles a standard digital mammogram, and a high-energy image. The high-energy image alone is not directly interpretable, but when mathematically combined with the low-energy image through a subtraction process, the result suppresses normal background breast tissue while revealing areas where the contrast agent has accumulated -- highlighting lesions based on their blood supply rather than just their density. This makes CESM significantly more sensitive than standard DM, particularly for lesions hidden within dense tissue.
Despite CESM's clinical advantages, AI development for breast imaging has been almost entirely based on digital mammography datasets. As of this study's publication, no public dataset of CESM images existed for training or validating AI systems. This means that AI models capable of interpreting CESM -- the more sensitive modality -- could not be developed or compared. The CDD-CESM (Categorized Digital Database for CESM) dataset introduced in this paper is the first publicly available CESM dataset, paired with low-energy images, expert annotations, segmentation masks, and full medical reports.
The dataset was collected at the Radiology Department of the National Cancer Institute, Cairo University, Egypt from January 2019 to February 2021. Institutional review board approval and patient informed consent were obtained from 326 female patients aged 18 to 90 years. Two mammography machines were used -- a GE Healthcare Senographe DS and a Hologic Selenia Dimensions system -- with all other data acquisition and post-processing parameters kept identical across both systems. This dual-machine inclusion reflects real-world clinical variation and makes the dataset more generalizable than single-vendor collections.
Each CESM examination produces both low-energy and subtracted CESM images in both CC (craniocaudal, top-down) and MLO (mediolateral oblique, angled) views, typically yielding eight images per patient. The dataset contains 2,006 images in total: 1,003 low-energy images paired with 1,003 subtracted CESM images. Forty-six patients had only four images due to mastectomy, and 87 patients had missing images due to quality concerns. All DICOM images were exported losslessly to JPEG format, with approximately 30% manually cropped to remove irrelevant borders. Images have an average resolution of 2355x1315 pixels -- clinical quality suitable for detecting fine findings.
A unique feature of the dataset is its comprehensive annotation structure. Each image was reviewed independently by two radiologists using the ACR BI-RADS 2013 lexicon, the international standard for mammography reporting. Annotations cover breast composition (density category A through D), mass characteristics (shape, margin, density), architectural distortion, asymmetry, calcification type and distribution, mass and non-mass enhancement patterns, overall BI-RADS assessment (categories 1 through 6), and final pathological diagnosis or follow-up outcome. Full written medical reports are included for every case.
The 2,006 images are distributed across a clinically realistic spectrum of findings. The largest category is normal examinations with 751 images (plus 6 post-neoadjuvant cases with no residual disease). Abnormal findings include 310 mass images, 334 mass enhancement images, 238 calcification images, 222 asymmetry images, 184 non-mass enhancement images, 159 postoperative images, 48 architectural distortion images, and 8 post-neoadjuvant chemotherapy images. A total of 248 images contain more than one finding simultaneously, reflecting the complexity of real clinical cases. The three-way pathological split is 37.4% normal, 29.3% benign, and 33.3% malignant.
The cancer type distribution among malignant cases reflects real-world breast cancer epidemiology. Invasive ductal carcinoma is the most common at 67.5% of cancer cases, followed by inflammatory breast cancer (6.3%), invasive lobular carcinoma (6.3%), mixed invasive ductal and lobular carcinoma (4.2%), purely ductal carcinoma in situ (2.5%), and other types (13.5%). The age distribution covers the full screening range: 17.8% of patients are under 40, 30.7% are in the 40-49 range, 29.1% are 50-59, 18.1% are 60-69, and 4.3% are 70 or older. This breadth of age representation is particularly valuable given that younger patients are often underrepresented in breast imaging AI datasets.
One practically important annotation is the machine label -- indicating which of the two mammography systems acquired each image. This enables researchers to study whether AI models trained on one vendor's equipment generalize to the other, a key question for clinical deployment. The dataset's diverse annotation system includes 140 unique standardized descriptor tags drawn directly from the BI-RADS lexicon, enabling fine-grained research into specific lesion characteristics rather than just broad malignant-versus-benign classification.
Manual segmentation -- drawing precise outlines around abnormal findings -- is the most time-consuming component of medical image annotation. To reduce this burden, the dataset is accompanied by an automated segmentation approach that requires no radiologist input beyond the classification labels already present in the dataset. The method uses EfficientNetB0, a compact and efficient convolutional neural network architecture pretrained on ImageNet, fine-tuned to classify each image as normal, benign, or malignant.
Once the classification model is trained, Grad-CAM (Gradient-weighted Class Activation Mapping) is applied to generate heatmaps that highlight which image regions most influenced the model's classification decision. Grad-CAM works by computing gradients of the predicted class score with respect to the final convolutional layer's feature maps, then weighting those maps by gradient magnitude to show where the model was most attentive. Importantly, the heatmaps are computed relative to the ground-truth class label (not the predicted class), ensuring the maps reflect the regions relevant to the correct diagnosis rather than whatever the model happened to predict.
The actual segmentation is extracted from the heatmaps using two thresholds applied sequentially. First, the top 25% of heatmap intensity values define candidate abnormal regions. Second, these regions are refined using the top 15% of white pixel intensities in the original image -- since abnormal findings often appear as bright (white) regions in contrast-enhanced images. Highlighting artifacts in image corners from normal pectoral muscle are explicitly excluded. This two-step thresholding approach converts the continuous heatmap into a discrete segmentation mask without requiring any additional labeled segmentation training data.
The automatic segmentation was validated against expert hand-drawn annotations using three complementary metrics: Intersection over Union (IOU), measuring overlap between predicted and hand-drawn regions; Dice coefficient (F1), a related overlap measure less sensitive to class imbalance; and Overlap50, the percentage of images where the automatic segmentation overlaps at least 50% of the hand-drawn annotation. Overall performance was IOU 64.2%, Dice 71%, and Overlap50 83.3% -- indicating the automated approach correctly localizes the general abnormal region in the large majority of cases.
Performance varied meaningfully across finding types in clinically interpretable ways. Mass enhancement achieved the highest Overlap50 at 91%, because enhancing masses are clearly visible and precisely localized in the subtracted CESM images. Postoperative cases had the lowest Overlap50 at 77%, as post-surgical edema and skin thickening create diffuse enhancement patterns that the model cannot distinguish from focal abnormalities. Malignant findings (Overlap50 90%) were segmented more accurately than benign findings (75%), largely because non-enhancing benign lesions are invisible or subtle in the subtracted images that the contrast-based approach is designed to highlight.
Subtracted CESM images (Overlap50 86%) were segmented more accurately than low-energy DM images (81%), confirming that suppression of normal background tissue in the subtracted images makes abnormalities easier for the AI to localize. Older patients (70+) had the highest Overlap50 at 94%, likely because older patients typically have less dense breasts, making lesions more conspicuous. Patients under 40 had the lowest accuracy (78%), consistent with the known relationship between breast density and detection difficulty. Performance was virtually identical between MLO and CC views, suggesting the model is robust to view orientation.
The CDD-CESM is the first dataset to make CESM images accessible for AI research, addressing a gap that has prevented the development and validation of AI systems for this clinically superior modality. Because CESM is more sensitive than standard digital mammography -- particularly in dense breasts -- an AI system trained on CESM data could be substantially more effective at finding cancers that standard DM AI would miss. The dataset is publicly available through The Cancer Imaging Archive repository, with helper code on GitHub for preprocessing, training, and evaluation.
The inclusion of both low-energy and subtracted CESM images for each patient is a deliberate design choice that mirrors clinical practice: radiologists routinely review both image types together. For AI development, this pairing allows researchers to build systems that combine information from both images, potentially capturing complementary information -- the structural detail of the low-energy image and the contrast-enhancement pattern of the subtracted image -- in a single multimodal model. The authors explicitly recommend using both image types together to maximize reliability.
The full-text medical reports included in the dataset open a research direction beyond image analysis alone: automated radiology report generation. Most large mammography datasets contain structured labels but not the narrative clinical reports that radiologists actually write. The CDD-CESM's reports follow the standardized BI-RADS lexicon, making them suitable for training natural language processing models that could assist in consistent, structured report writing -- a task with significant workflow implications given the volume of screening mammograms performed annually.