Abstract
Contrast-enhanced spectral mammography (CESM) is a relatively recent imaging modality with increased diagnostic accuracy compared to digital mammography (DM). New deep learning (DL) models were developed that have accuracies equal to that of an average radiologist. However, most studies trained the DL models on DM images as no datasets exist for CESM images. We aim to resolve this limitation by releasing a Categorized Digital Database for Low energy and Subtracted Contrast Enhanced Spectral Mammography images (CDD-CESM) to evaluate decision support systems. The dataset includes 2006 images, with an average resolution of 2355 × 1315, consisting of 310 mass images, 48 architectural distortion images, 222 asymmetry images, 238 calcifications images, 334 mass enhancement images, 184 non-mass enhancement images, 159 postoperative images, 8 post neoadjuvant chemotherapy images, and 751 normal images, with 248 images having more than one finding. This is the first dataset to incorporate data selection, segmentation annotation, medical reports, and pathological diagnosis for all cases. Moreover, we propose and evaluate a DL-based technique to automatically segment abnormal findings in images.
Measurement(s) | Dual-Energy Contrast-Enhanced Digital Spectral Mammography |
Technology Type(s) | digital curation |
Sample Characteristic - Organism | Homo sapiens • Breast |
Sample Characteristic - Location | Egypt |
Similar content being viewed by others
Background & Summary
Digital mammography (DM) is the gold standard imaging modality for early detection of breast cancer. However, limitations exist in patients with dense breasts as its overall sensitivity decreases1. Contrast-enhanced spectral mammography (CESM) is a contrast-based digital mammogram that has been approved by the Food and Drug Administration (FDA) in 2011 to be used as an adjunct to DM and ultrasound examinations for localization and characterization of occult or inconclusive lesions. Dual-energy image acquisition is performed where low and high-energy images are obtained. Several studies proved that low-energy images obtained appear like the standard DM images and are non-inferior to them2. High-energy images are non-interpretable; to overcome this, low and high-energy images are recombined and subtracted through appropriate image processing to suppress the background breast parenchyma after the acquisition. Figure 1 shows the resulting subtracted images obtained for interpretation, revealing contrast enhancement areas in a suppressed breast tissue background. Findings could be identified according to their density, morphologic, and enhancement characteristics3. However, estimating whether a lesion is benign or malignant without being seen by a radiologist is challenging due to the significant variation in the lesions’ visual characteristics4.
Computer-aided detection (CAD) systems were introduced in the early 2000’s to help radiologists interpret mammography images. However, this proved to be challenging in clinical practice due to the increased rate of false positives marked by the CAD systems, which can distract the radiologists5. Currently, the use of artificial intelligence (AI) in radiology is still in its early stages. Nonetheless, algorithms that analyze pixel data distinguish patterns from images that might not have been previously identified even by expert radiologists6. Deep learning (DL) has a promising potential in performing many tasks such as automatically detecting lesions and helping radiologists provide a more accurate diagnosis. Moreover, new multimodal DL models like the perceiver7 make it feasible to train on large datasets and extract good unsupervised image representations that can be used on a wide range of tasks. However, fully annotated and large-sized datasets are required and will be crucial for training new DL networks or fine-tuning existing pre-trained DL networks and evaluating them. This is why it is important for radiologists to understand the impact of these machine-learning (ML) based analytical tools and recognize how they might influence and change the radiological practice soon8.
In the past couple of years, a small number of public mammography datasets were released, including the Digital Database for Screening Mammography (DDSM)9, the Image Retrieval in Medical Applications (IRMA) project10, the Mammographic Imaging Analysis Society (MIAS) database11, and the Curated Breast Imaging Subset of DDSM (CBIS-DDSM)012. These datasets contain DM images only, and none include CESM images.
In this paper, we present a CESM categorized dataset that provides easily-accessible low energy images with corresponding subtracted CESM images, abnormality segmentation annotation, verified medical reports, and pathological diagnosis for all cases. It will add to the ongoing advancements in future mammography DL-based systems. We also propose a new DL-based technique to automatically segment the abnormal findings in the images without intervention from radiologists, as segmentation annotation is a time-consuming task.
Methods
We collected and reformatted the data into an easily-accessible format. Figure 2 displays the flow diagram of the process to prepare our dataset: image preprocessing, manual annotations, and the automatic segmentation.
Technique of contrast enhanced mammography examination
CESM is done using the standard DM equipment but with additional software that performs dual-energy image acquisition. Two minutes after intravenously injecting the patient with non-ionic low-osmolar iodinated contrast material (dose: 1.5 mL/kg), craniocaudal (CC) and mediolateral oblique (MLO) views are obtained. Each view comprises two exposures, one with low energy (peak kilo-voltage values ranging from 26 to 31kVp) and one with high energy (45 to 49 kVp). A complete examination is carried out in about 5–6 minutes.
Description of dataset
The dataset is a collection of low-energy images with their corresponding subtracted CESM images gathered from the Radiology Department of the National Cancer Institute, Cairo University, Egypt over the period from January 2019 to February 2021. The images are all high resolution with an average of 2355 × 1315 pixels. Institutional review board approval and patient informed consent to carry out and publish data were obtained from 326 female patients aged from 18 to 90 years. The dataset contains 2006 images with CC and MLO views (1003 low energy images and 1003 subtracted CESM images), samples of low energy and subtracted CESM images are shown in Fig. 3. Usually, each patient has a total of 8 images, 4 images for each breast side consisting of low energy and subtracted CESM images for each CC and MLO view. However, there are 46 patients with only 4 images as they had mastectomy on a breast side, and 87 patients with missing images as some were not available or removed due to quality concerns. Two different machines were used for image acquisition; GE Healthcare Senographe DS and Hologic Selenia Dimensions Mammography Systems. The two machines provide similar quality, and all other steps in the data acquisition and post-processing phases were kept the same. The images are manually-annotated by expert radiologists according to the American College of Radiology Breast Imaging Reporting and Data System (ACR BIRADS) 2013 lexicon for standardized descriptors13. The annotations, shown in Table 1, include breast composition, mass shape, mass margin, mass density, architectural distortion, asymmetries, calcification type, calcification distribution, mass enhancement pattern, non-mass enhancement pattern, non-mass enhancement distribution, and overall BIRADS assessment (1 to 6). Both follow-up and pathological results are also included in the annotations, as pathological results are the gold-standard reference for radiologically-suspicious or malignant-looking lesions, and follow-up is the gold standard for benign-looking lesions. Moreover, full medical reports, written by an ensemble of radiologists, are provided for each case along with manual segmentation annotation for the abnormal findings in each image.
Annotations
Data are gathered and stored in a DICOM format. Some irrelevant annotations that are not used for lesion identification and classification were removed, including the patient’s name, ID, date of the study, and the image series. Each image with its corresponding annotation was compiled into one comma-separated-value (CSV) file.
Medical reports
Separate corresponding reports for the CESM images and the DM images are also included in the dataset. Each report consists of the findings, depicted for each breast side separately, written following the ACR BIRADS 2013 lexicon for standardized descriptors and reporting associated with the BIRADS category annotated for the case. All patients’ identification data were removed. We believe that releasing the full-text medical reports is important, as research studies concerned with radiology report-writing often struggle with the lack of full reports not being present in large datasets14.
Image processing
DICOM images were exported losslessly to a joint photographic experts group (JPEG) format using RadiAnt DICOM viewer application(https://www.radiantviewer.com/). After automatically removing all irrelevant data from each image, around 30% of the images were manually cropped to eliminate all unused and irrelevant boundaries. Furthermore, the images are named as follows {patient number}_{breast side}_{image type}_{image view}; example ‘P1_L_CM_MLO’.
Segmentation visual model
In this section, we describe our method to automatically segment the abnormal parts of the images. A deep learning model, EfficientNetB0, was trained to predict the overall diagnosis (Normal, Benign, Malignant). GradCam15 was used to generate highlights for the parts of the image that contributed to the model’s prediction. A threshold of the top 25% GradCam intensities is then used on the highlights to generate the segments. Furthermore, a threshold of the top 15% white pixels is used to further finetune the segmentations.
Preprocessing
The images were first resized to be 224 × 224 using interpolation and anti-aliasing. Then the images were normalized by subtracting from the mean and dividing by the standard deviation. Random image augmentations were also used like cropping, zooming, and horizontal flipping. Furthermore, we experimented with non-traditional data augmentation methods16 which uses generative adversarial networks (GANs) to generate new images. However, the generated images did not satisfy the experts, so only traditional data augmentations were used.
Model & training
An EfficientNetB017, pre-trained on ImageNet18, was used as the starting model in our experiments. We finetuned the model by removing the final layer and adding a layer with three output classes (Normal, Benign, Malignant). All the weights are left to be fine-tuned during the training. Categorical cross-entropy was used as the loss function with Adam optimizer19 as shown in Eq. 1, where CE(b) is the cross entropy loss for batch b, C the number of classes, N the number of images in the batch, y is the ground-truth, and \(\widehat{y}\) is the prediction. A batch size of 16 was used, a decaying learning rate of 1e-3, and a dropout layer20 with a drop probability of 0.8 on the final visual features was used before the classifier.
Highlights
After the model achieved a good accuracy on all the images, we used GradCam15 to get heatmaps representing the parts of the image that had the highest impact on the model’s decision. The heatmaps are traced back from the ground-truth class and not the predicted class. Moreover, we removed any highlights in the corners of the image as they are often present at the location of normal pectoral muscles.
Segmentation
To get the actual pixel segmentation, we used the top 25% of the heatmap’s intensities to serve as the abnormal segment. Moreover, to finetune the segments on the exact abnormality, we used the intersection of the segments and the top 15% white pixel intensities of the image as shown in Fig. 4.
Data Records
The low energy and subtracted CESM images are distributed as JPEG files. They include both MLO and CC views of the mammograms.
Metadata for each image is incorporated as an associated CSV file consisting of:
-
Path to image files
-
Patient number
-
Breast side: Left or Right
-
Type of Examination: DM (low energy image) or CESM (subtracted image)
-
View: CC or MLO
-
Density category (if low energy image)
-
Number of findings (if multiple)
-
Mass shape, density, and margin (if present)
-
Mass enhancement pattern (if present)
-
Architectural distortion (if present)
-
Asymmetry (if present)
-
Calcification type and distribution (if present)
-
Non-mass enhancement pattern and distribution (if present)
-
BI-RADS assessment
-
Pathology: Benign or Malignant
Figure 5 shows histograms of BIRADS category and the corresponding final pathology/follow up result. Table 2 displays the characteristics of the CDD-CESM dataset.
The CDD-CESM dataset is available21 on The Cancer Imaging Archive repository22. The dataset includes all images, annotations, and full medical reports.
Technical Validation
For the segmentation evaluation of our DL model, experienced radiologist provided hand-drawn segmentations for each abnormal finding in the CDD-CESM dataset. We calculated the intersection over union (IOU) and the dice coefficients (F1) between the computed and hand-drawn segmentations, after applying the same white-intensity threshold on the hand-drawn segmentations. Furthermore, we added another metric which we called overlap50, which is the percentage of images where the automatic segmentation overlaps with at least 50% of hand-drawn segmentation. The average IOU was 64.2% overall, overlap50 was 83.3%, and the average F1 was 71% overall. We also calculated these metrics separately for different groups of images according to the following criteria:
Different findings represented in the dataset
Mass enhancement had the highest overlap50 = 91%. Furthermore, postoperative cases had the lowest overlap50 = 77%. This might be attributed to post operative edematous changes and skin thickening that are not accurately or completely observed by our DL model.
Age of patient
Patients aged seventy years and higher had the highest overlap50 = 94%. Forty years and lower had the lowest overlap50 = 78%. As expected, the accuracy of visualization decreases as the breast density increases.
Low energy or subtracted image
Low energy image overlap50 = 81%, compared to 86% in subtracted images. This might be due to the dense adenotic tissue in low-energy images obscuring abnormalities found behind it, which are suppressed in subtracted images. Thus, we recommend that radiologists use both low energy and subtracted images for each patient in each view, to increase reliability of using our DL technique in drawing their final conclusions.
Mediolateral or Craniocaudal view
We found the results to be comparable without much difference in terms of automatic segmentation output.
Benign or malignant finding
Benign findings had the lower overlap50 = 75% compared to 90% for malignant findings. Most of the benign lesions were non-enhancing in subtracted images. Furthermore, in low-energy images, benign lesions were either hidden behind the dense breast tissues, had equal density or parallel orientation to the surrounding breast parenchyma. However, highly cellular benign findings were accurately depicted by our DL model. Decreased accuracy was found with multiplicity and retroareolar locations.
Generally, decreased accuracy of detection by our DL model was also present in some subtracted images with halo (breast-within-breast) or ripple artifacts. These calculations are shown in Table 3, and example outputs from our DL model are showed in Fig. 6.
Usage Notes
The dataset can be used to train machine learning models to classify mammogram images into normal, benign, and malignant, or classify the tags associated with each image. Moreover, it can be used to train segmentation models to segment the lesions. Furthermore, the full-text medical reports can be used to train report generation models.
Code availability
A Github repository is publicly available (https://github.com/omar-mohamed/CDD-CESM-Dataset) which contains helper scripts to make training a DL model on the dataset easier like reading the annotations, pre-processing the images by resizing and normalizing, training different existing models, augmenting the images while training, and evaluating the different models and plotting the segmentation results. The scripts were written using Python 3.6 with Tensorflow 2.3 for the training process, and OpenCV 4.1 and Pillow 6.1 for the image processing.
References
Chetlen, A., Mack, J. & Chan, T. Breast cancer screening controversies: who, when, why, and how? Clinical imaging. 40, 279–282 (2016).
Lalji, U. et al. Evaluation of low-energy contrast-enhanced spectral mammography images by comparing them to full-field digital mammography using euref image quality criteria. European radiology. 25, 2813–2820 (2015).
Bhimani, C. et al. Contrast-enhanced spectral mammography: technique, indications, and clinical applications. Academic radiology. 24, 84–88 (2017).
Lewin, J. M., Isaacs, P. K., Vance, V. & Larke, F. J. Dual-energy contrast-enhanced digital subtraction mammography: feasibility. Radiology. 229, 261–268 (2003).
Hupse, R. et al. Computer-aided detection of masses at mammography: interactive decision support versus prompts. Radiology. 266, 123–129 (2013).
Erickson, B. J. Machine learning: discovering the future of medical imaging. J Digit Imaging. 30, 391 (2017).
Jaegle, A. et al. Perceiver: General perception with iterative attention (2021).
Lee, J.-G. et al. Deep learning in medical imaging: general overview. Korean journal of radiology 18, 570 (2017).
Heath, M., Bowyer, K., Kopans, D., Moore, R. & Kegelmeyer, W. P. The digital database for screening mammography. In Proceedings of the Fifth International Workshop on Digital Mammography, 212–218, http://www.eng.usf.edu/cvprg/Mammography/Database.html (2001).
Lehmann, T. M. et al. Content-based image retrieval in medical applications. Methods of information in medicine 43, 354–361 (2004).
Suckling J, P. The mammographic image analysis society digital mammogram database. Digital Mammo 375–386 (1994).
Lee, R. S. et al. A curated mammography data set for use in computer-aided detection and diagnosis research. Scientific data 4, 1–9 (2017).
D’Orsi, C. 2013 ACR BI-RADS Atlas: Breast Imaging Reporting and Data System 5th edn (American College of Radiology, 2014).
Alfarghaly, O., Khaled, R., Elkorany, A., Helal, M. & Fahmy, A. Automated radiology report generation using conditioned transformers. Informatics in Medicine Unlocked 24, 100557 (2021).
Selvaraju, R. R. et al. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis. 128, 336–359 (2020).
Al-Dhabyani, W., Gomaa, M., Khaled, H. & Fahmy, A. Deep learning approaches for data augmentation and classification of breast masses using ultrasound images. International Journal of Advanced Computer Science and Applications 10 (2019).
Tan, M. & Le, Q. V. Efficientnet: Rethinking model scaling for convolutional neural networks. Proceedings of the 36th International Conference on Machine Learning 97, 6105–6114 (2019).
Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems 25, 1106–1114 (2012).
Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. In Bengio, Y. & LeCun, Y. (eds.) 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings (2015).
Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014).
Khaled, R. et al. Categorized digital database for low energy and subtracted contrast enhanced spectral mammography images. The Cancer Imaging Archive https://doi.org/10.7937/29kw-ae92 (2021).
Clark, K. W. et al. The cancer imaging archive (TCIA): maintaining and operating a public information repository. J. Digit. Imaging 26, 1045–1057, https://doi.org/10.1007/s10278-013-9622-7 (2013).
Funding
Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooperation with The Egyptian Knowledge Bank (EKB).
Author information
Authors and Affiliations
Contributions
M.H. directed the project, examined the images, assigned BIRADS category for each case, and reviewed the paper. R.K. examined the images, written the medical reports, and manually provided the hand-drawn segmentation annotations for each case. R.K. also participated in writing the paper and organizing the metadata. O.M. supported the study, edited and reviewed the paper. H.K. supported the study, participated in writing and editing of the paper. O.A. was responsible for training the segmentation model, writing the code for data handling, and participated in writing the paper. A.E. and A.F. supervised the training of the segmentation model, and all the technical details.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Khaled, R., Helal, M., Alfarghaly, O. et al. Categorized contrast enhanced mammography dataset for diagnostic and artificial intelligence research. Sci Data 9, 122 (2022). https://doi.org/10.1038/s41597-022-01238-0
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41597-022-01238-0
This article is cited by
-
YOLO-based CAD framework with ViT transformer for breast mass detection and classification in CESM and FFDM images
Neural Computing and Applications (2024)
-
Determining the Differentiation of Benign and Malignant NME Lesions in Contrast-Enhanced Spectral Mammography Images Based on Convolutional Neural Networks
Journal of Medical and Biological Engineering (2023)
-
Breast cancer diagnosis from contrast-enhanced mammography using multi-feature fusion neural network
European Radiology (2023)