TN-Mammo: A Multi-view Mammography Dataset for Breast Density Classification

Abstract

Breast cancer is one of the most common types of cancer among women, leading to a growing and essential need for early and precise detection. A variety of machine learning techniques have been demonstrating great promise in improving diagnostic accuracy integrated with digital mammography, which remains the gold standard screening technique in the detection of early breast cancer. Since the performance of machine learning models generally depends on the quality of training data, it is essential to build a high-quality mammogram dataset. Although there exists a number of public datasets, there are still significant limitations due to both their quality and quantity. We present an extensive Vietnamese mammogram dataset with breast density annotations in an effort to bridge the gap between availability and practical usability. Our mammogram dataset named TNMammo, consists of bilateral craniocaudal (CC) and mediolateral oblique (MLO) views for 676 subjects, each with paired left and right breast views. The breast density and orientation for each case have been independently assessed by two radiologists in a double-blind manner, ensuring consistency and reliability in the annotations. Each case is evaluated based on breast density, categorized into four levels: A, B, C, and D.

Background

Breast cancer is the second most common cancer diagnosed in women, exceeded only by nonmelanoma skin cancer. It is the leading cause of death from cancer in women worldwide [1]. According to the GLOBOCAN [2] in 2022, female breast cancer accounted for 11.6% globally with 2.308.897 new cases reported, making it the second most common type of cancer among 33 distinct types, following lung cancer. Considering the mortality rate, breast cancer is responsible for 665,684 deaths, placing it the fourth. Particularly, the annual national report of Vietnam in 2022 identifies 24,563 new cases and 10,008 deaths, which are the first and the fourth, respectively [3]. To fight against this disease, scientists have developed a variety of methods to assist doctors with second opinions in diagnosing and treating patients.

Despite significant improvements in the diagnosis, detection, and treatment of breast cancer, its mortality rate has shown no signs of decreasing over the years, especially in developing countries [4]. The main reason is that patients often pay little attention to proactively examining symptoms of early-stage breast cancer. This leads to the disease being discovered too late, leading to the tumor having grown large or metastasized to other organs. This makes successful treatment for patients more difficult, thus significantly reducing patient survival rates when detected late. Previous studies have shown that patients who detect breast cancer early and have appropriate long-term treatment methods will have a higher survival rate [5, 6].

Traditional diagnostic methods such as mammography, ultrasound, MRI, and biopsy have played an important role in detecting the disease. However, each method has certain limitations. For example, in cases where patients have dense breast tissue or when tumor detection by mammography is difficult, using ultrasound will make diagnosis easier [7]. Conversely, ultrasound cannot detect microcalcifications. Therefore, mammography is indispensable in breast cancer screening [7]. However, the sensitivity of mammography is also significantly reduced in patients with dense breast tissue and especially in older patients [8]. Although MRI is highly effective in detecting breast cancer, it is often expensive to deploy and difficult to make widely available as a breast cancer screening method, particularly in regions with limited medical infrastructure [9–13].

In addition, the application of deep learning in medical imaging prediction, especially in breast cancer detection, relies heavily on high-quality datasets that possess key attributes: diverse image distribution, comprehensive annotations, and clinically validated ground truth. The remarkable advances in artificial intelligence (AI) have revolutionized breast cancer surveillance. However, this effect can only be achieved with standardized and scientifically managed mammography datasets. Over the past 30 years, as in Figure 1, the global research community has continuously developed and perfected deep learning image database systems, serving both the goals of model training and evaluating computer-aided detection (CAD) algorithms. The Mammographic Image Analysis Society released the MIAS [14] database of digitized film mammograms in 1994, containing 322 images - 161 cases from the UK’s national screening program. Each MIAS image was low-resolution (re-sampled to 1024×1024) with radiologist-marked abnormality locations. While groundbreaking at the time, the MIAS dataset’s limited size and image quality made it less suitable for training modern deep learning models which inherently need a lot of data. Subsequently, the Digital Database for Screening Mammography (DDSM) [15] was developed as a collaborative effort in the US (Massachusetts General Hospital, University of South Florida, etc.) and publicly published in 1997, which contains approximately 2,620 screening studies, each study includes four views, totaling 10,000 images from film mammograms with verified pathology labels. In 2012, Portugal released the INBreast [16] as one of the first public full-field digital mammography (FFDM) datasets, which contains 115 cases - 410 images with diverse findings. Though this dataset had a small size, it also set a new standard for annotation quality and ground-truth quality in public datasets. In the next few years, researchers created much larger mammography datasets from national screening programs and multi-center cohorts. For instance, the OPTIMAM Mammography Image Database (OMI-DB) [17] in the UK had grown into one of the largest repositories, including nearly 7 million digital mammogram images from about 465,000 women, which the National Health Service (NHS) had gathered over 10+ years of screening across multiple sites. Another example is the CSAW dataset (Cohort of Screen-Aged Women) [18] from Sweden, comprising all women ages 40–74 who participated in Stockholm regional screening between 2008–2015, recording 499,807 women with 1,182,733 screening exams. At the same scale, the NYU Breast Cancer Screening Dataset - NYU Breast [19] contains 1,001,093 mammograms from 229,426 screening exams, involving 141,472 patients, which were collected by staff at NYU Langone Medical Center. In Asia, researchers have expanded the geographic diversity of available mammography data with new public datasets. The Cancer Imaging Archive released the Chinese Mammography Database (CMMD) [20], which provides 3,728 FFDM images from 1,775 patients in China. Similarly, the VinDr-Mammo [21] dataset from Vietnam serves as a recent large-scale FFDM benchmark that includes 5,000 four-view mammography exams - 20,000 images acquired from Vietnamese hospitals.

A number of public mammogram datasets are available and have contributed significantly to the development of deep-learning models. Simultaneously, scientists have researched and developed various methods to analyze and advance breast cancer diagnosis. In 2022, Busaleh et al. [22] using the DDSM and Inbreast datasets, leveraging the information from the CC and MLO views, with methods such as breast segmentation and contrast enhancement to improve data quality before classification. Studies have also focused on the benign or malignant, normal or abnormal classification. Houssein et al. [23] proposed an improved marine predators algorithm IMPA-ResNet50 model that optimized the Resnet50 architecture and achieved 98.32% accuracy on the CBIS-DDSM dataset [24] and 98.88% on the MIAS dataset. Among them, convolutional neural networks (CNNs) have emerged as powerful models. Li et al. [25] integrating attention mechanisms into residual learning for breast density classification on both private and INbreast datasets. Beyond CNNs, other classification models, such as BASCNet (Zhao et al. [26]), include, which employs bilateral spatial and channel attention to reach 90.51% accuracy on the INbreast and DDSM datasets.

Researchers have also explored alternative approaches, such as Darweesh et al. [27], who used genetic programming to automatically construct classification attributes with mammographic data from a public repository. Additionally, they have developed hybrid methods that combine handcrafted features with traditional machine learning techniques like SVM, providing diverse strategies to detect and assess breast cancer risk across various datasets.

In this work, we have collected mammograms from Thong Nhat Hospital in Viet Nam for a wide range of ages and severities. We hope that this dataset will be of great help to the mammography research community, as well as the AI field, including the development of deep learning models for disease diagnosis, improving the efficiency of early breast cancer detection. We have performed some existing experiments on this dataset. In addition, we have performed some normalization techniques. The technical validation of this dataset, described further below, serves as proof of the dataset’s quality.

Methods

The study included women undergoing mammography-based breast cancer screening at Thong Nhat Hospital from January 2022 to June 2024. Digital mammograms were downloaded from the hospital’s PACS (Picture Archiving and Communication System) system to a workstation in the hospital to anonymize information and selected based on quality criteria, excluding images with defects, motion, or technical errors and women with previous breast cancer, surgery, or pregnancy. We used convenience sampling, and two radiologists independently assessed breast density and orientation in a double-blind manner. Discrepancies in their readings were resolved through consultation with the department head, and consensus results were used to train the model. We converted anonymized DICOM images to JPEG for model training because JPEG files are lighter while still preserving image quality [28]. This conversion reduces storage and computational requirements, making data processing more efficient. Despite compression, minimally compressed JPEG images retain essential anatomical details, allowing the model to detect subtle patterns in mammograms and achieve high classification accuracy.

Data Description

The project directory contains annotation files, namely TNMammo_labels.csv, and a subfolder images that contains JPEG files.

• images: contains 676 subdirectories corresponding to 676 exams in the dataset, where the folder name is an encrypted patient ([ID]). Each folder includes four JPEG files corresponding to the four standard mammogram views of each breast.
  
  The path to each image file is structured as: images/[id]/[view].jpg where [ID] is the encrypted patient, and [view] indicates the mammogram view type (e.g., left_cc, left_mlo, right_cc, right_mlo).

• TNMammo_labels.csv: Each row corresponds to a case and provides the breast density assessment. The attributes in each row are:
  • ID: The encrypted patient or case identifier.
  • Labels: The breast density category of patients. The possible values for the Labels attribute are A, B, C, and D.
  • Age: Age of the patients. The ages of patients in the dataset range from 18 to 91 years.

The dataset should be organized as follows:

datasets/
└── TNMammo/
    ├── images/
    │   └── ID/
    │       ├── left_cc.jpg
    │       ├── left_mlo.jpg
    │       ├── right_cc.jpg
    │       └── right_mlo.jpg
    └── TNMammo_labels.csv

Usage Notes

The dataset can be utilized for various machine learning and medical imaging applications, particularly in classification and diagnostic model development. Below are key considerations:

• Preprocessing Recommendations: Images may require normalization and resizing to fit specific model input requirements.
• Potential Applications:
  • Supervised Learning: The dataset is well-suited for classification tasks where models learn to predict labels based on image features.
  • Feature Extraction: Researchers can extract and analyze specific features related to breast density.
  • Multi-View Analysis: Since each record contains four views, multi-view fusion techniques can be explored for improved accuracy.
• Considerations:
  • Ensure proper data splitting (e.g., train/validation/test sets) to avoid data leakage.
  • The dataset should be used responsibly in accordance with ethical and legal guidelines, especially for medical applications.
• Known Limitations:
  • Dataset size: With only 676 exams, the dataset may limit the performance of deep learning models that require large-scale data.
  • Label scope: The dataset provides breast density categories (A–D) only, which restricts its use for lesion detection or abnormality classification.
  • Modality limitation: The dataset contains mammography images only, without complementary modalities such as ultrasound or MRI.
  • Population diversity: Images are collected from a single source, which may reduce demographic or device diversity and affect model generalization.

The dataset should be used responsibly in accordance with ethical and legal guidelines, especially for medical applications. All the scripts we used for the experiment are publicly available at [29].

Release Notes

This is the first public release of the TNMammo dataset.

Ethics

The dataset has received ethical approval from the Thong Nhat Hospital Ethics Committee (Certificate of Approval No. 43/2024/CN-BVTN-HDDD, signed in Ho Chi Minh City on 04 March 2024). We confirm that the dataset complies with all relevant local regulations governing data collection and sharing.

Acknowledgements

We would like to thank Thong Nhat Hospital for providing the mammography data used in this research. We would love to thank AISIA Research Lab, University of Science, Vietnam National University in Ho Chi Minh City, and Tan Tao University for supporting us in experiments and research algorithms for this project.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

1. Trayes KP, Cokenakes SE. Breast cancer treatment. Am Fam Physician. 2021;104(2):171-8.
2. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74(3):229-63.
3. Campbell SL, Gear CW. The index of general nonlinear DAEs. Numer Math. 1995;72(2):173-96.
4. Azamjah N, Soltan-Zadeh Y, Zayeri F. Global trend of breast cancer mortality rate: a 25-year study. Asian Pac J Cancer Prev. 2019;20(7):2015.
5. Migowski A. Early detection of breast cancer and the interpretation of results of survival studies. Cien Saude Colet. 2015;20(4):1309-10.
6. Wang L. Early diagnosis of breast cancer. Sensors. 2017;17(7):1572.
7. Gordon PB. Ultrasound for breast cancer screening and staging. Radiol Clin North Am. 2002;40(3):431-41.
8. Kolb TM, Lichy J, Newhouse JH. Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825 patient evaluations. Radiology. 2002;225(1):165-75.
9. Boetes C, Barentsz JO, Mus RD, van der Sluis R, van Erning L, Hendriks J, Holland R, Ruys S. MR characterization of suspicious breast lesions with a gadolinium-enhanced turboFLASH subtraction technique. Radiology. 1994;193(3):777-81.
10. Gilles R, Guinebretiere JM, Lucidarme O, Cluzel P, Janaud G, Finet JF, Tardivon A, Masselot J, Vanel D. Nonpalpable breast tumors: diagnosis with contrast-enhanced subtraction dynamic MR imaging. Radiology. 1994;191(3):625-9.
11. Harms SE, Flamig DP, Hesley KL, Meiches MD, Jensen RA, Evans WP, Savino DA, Wells RV. MR imaging of the breast with rotating delivery of excitation off resonance: clinical experience with pathologic correlation. Radiology. 1993;187(2):493-501.
12. Morris EA, Schwartz LH, Dershaw DD, Van Zee K, Abramson AF, Liberman L. MR imaging of the breast in patients with occult primary breast carcinoma. Radiology. 1997;205(2):437-40.
13. Orel SG, Schnall MD, LiVolsi VA, Troupin RH. Suspicious breast lesions: MR imaging with radiologic-pathologic correlation. Radiology. 1994;190(2):485-93.
14. Suckling J. The Mammographic Image Analysis Society digital mammogram database. In: Digital Mammography. Dordrecht: Springer; 1994. p. 375-86.
15. Heath M, Bowyer KW, Kopans D, Moore R, Kegelmeyer WP. The Digital Database for Screening Mammography. In: Yaffe MJ, editor. Proceedings of the 5th International Workshop on Digital Mammography. Madison: Med Phys Publ; 2000. p. 212-8.
16. Moreira IC, Amaral I, Domingues I, Cardoso A, Cardoso MJ, Cardoso JS. INbreast: toward a full-field digital mammographic database. Acad Radiol. 2012;19(2):236-48. doi:10.1016/j.acra.2011.09.014.
17. Halling-Brown MD, Warren LM, Ward D, Lewis E, Mackenzie A, Wallis MG, Wilkinson LS, Given-Wilson RM, McAvinchey R, Young KC. OPTIMAM Mammography Image Database: a large-scale resource of mammography images and clinical data. Radiol Artif Intell. 2020;3(1):e200103.
18. Dembrower K, Lindholm P, Strand F. A multi-million mammography image dataset and population-based screening cohort for the training and evaluation of deep neural networks—the Cohort of Screen-Aged Women (CSAW). J Digit Imaging. 2020;33(2):408-13. doi:10.1007/s10278-019-00278-0.
19. Wu N, Phang J, Park J, Shen Y, Kim S, Heacock L, Moy L, Cho K, Geras K. The NYU Breast Cancer Screening Dataset v1.0. New York: New York Univ.; 2019. Technical Report.
20. Cai H, Wang J, Dan T, Li J, Fan Z, Yi W, Cui C, Jiang X, Li L. An online mammography database with biopsy confirmed types. Sci Data. 2023;10(1):123.
21. Nguyen HT, Nguyen HQ, Pham HH, Lam K, Le LT, Dao M, Vu V. VinDr-Mammo: a large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Sci Data. 2023;10(1):277.
22. Busaleh M, Hussain M, Aboalsamh HA, Al Sultan SA. TwoViewDensityNet: two-view mammographic breast density classification based on deep convolutional neural network. Mathematics. 2022;10(23):4610.
23. Houssein EH, Emam MM, Ali AA. An optimized deep learning architecture for breast cancer diagnosis based on improved marine predators algorithm. Neural Comput Appl. 2022;34(20):18015-33.
24. Lee RS, Gimenez F, Hoogi A, Miyake KK, Gorovoy M, Rubin DL. A curated mammography dataset for use in computer-aided detection and diagnosis research. Sci Data. 2017;4:170177. doi:10.1038/sdata.2017.177.
25. Li C, Xu J, Liu Q, Zhou Y, Mou L, Pu Z, Xia Y, Zheng H, Wang S. Multi-view mammographic density classification by dilated and attention-guided residual learning. IEEE/ACM Trans Comput Biol Bioinform. 2020;18(3):1003-13.
26. Zhao W, Wang R, Qi Y, Lou M, Wang Y, Yang Y, Deng X, Ma Y. BASCNet: bilateral adaptive spatial and channel attention network for breast density classification in the mammogram. Biomed Signal Process Control. 2021;70:103073.
27. Ain QU, Xue B, Al-Sahaf H, Zhang M. A genetic programming approach to automatically construct informative attributes for mammographic density classification. In: 2022 IEEE International Conference on Data Mining Workshops (ICDMW); 2022 Nov 28-Dec 1; Orlando, FL. Piscataway, NJ: IEEE; 2022. p. 378-87.
28. Trieu PDY, Barron M, Lewis SJ. Use of full-quality DICOM images compared to minimally compressed mammograms in JPEG format for radiology training: a study from radiologist and radiographer perspectives. Acad Radiol. 2023;30(8):1748-55. doi:10.1016/j.acra.2022.11.012.
29. TNMammo GitHub Repository. Available from: https://github.com/LeHongCat/TNMammo [Accessed 2025 Sep 12].