~/yuki-kashiwada

$ cat research/tompei-cmmd.md

TOMPEI-CMMD Dataset

DatasetFirst creatorLicense: CC BY 4.0

Creators

  1. Yuki Kashiwada,
  2. Eichi Takaya,
  3. Mei Hiroya,
  4. Nanako Matsuda,
  5. Takumi Yashima,
  6. Tomoya Kobayashi,
  7. Gen Tamiya,
  8. Takuya Ueda

The Cancer Imaging Archive · Version 1 · https://doi.org/10.7937/wezw-bh22

  • DOI registered:
  • Version 1 updated:

TL;DR

TOMPEI-CMMD adds polygon lesion segmentations (mass, calcification, and four other finding types), lesion locations, corrected laterality/view labels, and clinical data to the Chinese Mammography Database (CMMD). It contains 1,773 masks for 1,385 breasts from 1,363 patients and is published on TCIA under CC BY 4.0.

##Key numbers

Patients
1,363
Breasts with lesions (MLO)
1,385
Segmentation masks
1,773
255 benign / 1,515 malignant / 3 excluded
Finding categories
6
Source images (DICOM)
4,162
18.29 GB
License
CC BY 4.0

##Background

The original CMMD on TCIA contains 3,728 mammograms from 1,775 patients examined between July 2012 and January 2016, with biopsy-confirmed benign/malignant diagnoses (and molecular subtypes for some patients). It has no lesion segmentations, so it gives no information on tumor location, size, shape, or margins, which makes it hard to use for training lesion-detection models.

##Data

From the CMMD images (acquired on a GE Senographe DS), left and right breasts were treated as separate images and only MLO views were kept, removing 1,127 CC views. This gave 2,601 breast images. After excluding 140 images with mammographically occult cancers and 25 with insufficient image quality, 2,436 images remain in the final clinical spreadsheet.

Segmentations cover the 1,385 breasts with lesions (1,363 patients), with 1,773 masks in total: 255 benign, 1,515 malignant, and 3 later excluded from the published paper. Lesion-free breasts are not included in the segmentation files.

##Annotation

A board-certified radiologist with 20 years of breast-imaging experience read all 2,601 images, with access to the CMMD benign/malignant outcomes, and recorded each finding and its location. Five radiological technologists then drew the segmentations under the radiologist's guidance.

Annotations are polygon segmentation masks, not bounding boxes, stored as JSON with the fields _id, type, color, cgPoints (contour coordinates), and x/y (lesion location). There are six finding tags: mass (blue), calcification (yellow), fibroadenoma (red), distortion (green), focal asymmetric density (purple), and lipoma (light blue). Images with several findings carry several tags.

##What's included

The release contains the lesion segmentations (JSON in a ZIP, 24.8 MB, 1,385 files), clinical data (XLSX, 624 KB, covering classification, demographics, diagnosis, and pathology), and a manifest for the source images. The source images (4,162 CMMD DICOM images, 18.29 GB) are not redistributed and are downloaded with the TCIA Data Retriever.

##Significance

Compared with CMMD, it adds (1) lesion segmentations with supporting clinical data, (2) corrections of left/right and CC/MLO labelling errors, and (3) lesion locations where the original had only per-image labels. It supports the development and validation of deep-learning breast cancer CAD systems and was used to train and evaluate MIME-ViT.

##Limitations

  • -MLO views only; there are no CC-view segmentations.
  • -Lesion-free breasts are not included in the segmentation files.
  • -The source data come from a single patient population (Chinese patients) and a single scanner model (GE Senographe DS).
  • -Source images are not bundled and must be downloaded separately from CMMD.

##How to download

  1. Open the TCIA dataset page (doi.org/10.7937/WEZW-BH22).
  2. Download the segmentations (TOMPEI-CMMD_v01_20250123.zip, 24.8 MB) and the clinical data (TOMPEI-CMMD_clinical_data_v01_20250121.xlsx, 624 KB) directly.
  3. For the source images, download the manifest (TOMPEI-CMMD_SourceImages_v01.tcia) and open it in the TCIA Data Retriever to fetch the CMMD DICOM images (4,162 images, 18.29 GB). The source collection is CMMD (DOI 10.7937/tcia.eqde-4b16).
  4. Use the sample Python code on GitHub (javasparrows/TOMPEI-CMMD) to view the masks overlaid on the images.
  5. Follow the TCIA Data Usage Policy and include the required TCIA data citation in your publications.

##FAQ

Q.What is the TOMPEI-CMMD dataset?

It is an extension of the public CMMD mammography collection that adds 1,773 lesion segmentations, lesion locations, and clinical data for 1,385 MLO-view breast images from 1,363 patients. It was published by Yuki Kashiwada et al. on The Cancer Imaging Archive (TCIA), DOI 10.7937/WEZW-BH22.

Q.How can I download TOMPEI-CMMD?

Download the segmentations (ZIP) and clinical data (XLSX) directly from the TCIA dataset page. To get the source mammograms (DICOM), open the .tcia manifest from the same page in the TCIA Data Retriever. See the download section on this page for step-by-step instructions.

Q.What annotations does TOMPEI-CMMD include?

Polygon segmentation masks in JSON, not bounding boxes. Each mask has one of six tags (mass, calcification, fibroadenoma, distortion, focal asymmetric density, lipoma) and lesion location coordinates.

Q.How is TOMPEI-CMMD different from CMMD?

CMMD only has per-image benign/malignant labels. TOMPEI-CMMD adds lesion segmentations, lesion locations, and clinical data, and corrects left/right and CC/MLO labelling errors.

Q.What is the license of TOMPEI-CMMD?

The TOMPEI-CMMD files are released under CC BY 4.0, so they can be reused with attribution. Users must follow the TCIA Data Usage Policy and include the required data citation.

Q.How do I cite TOMPEI-CMMD?

Use the TCIA data citation, reproduced verbatim in the citation section of this page. If you use the source images, also cite the CMMD collection (DOI 10.7937/tcia.eqde-4b16).

##How to cite

TCIA data citation (required)

Kashiwada, Y., Takaya, E., Hiroya, M., Matsuda, N., Yashima, T., Kobayashi, T., Tamiya, G., & Ueda, T. (2025). TOMPEI-CMMD Dataset (Version 1) [Dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/WEZW-BH22

BibTeX

@misc{kashiwada2025tompeicmmd,
  author       = {Kashiwada, Yuki and Takaya, Eichi and Hiroya, Mei and Matsuda, Nanako and Yashima, Takumi and Kobayashi, Tomoya and Tamiya, Gen and Ueda, Takuya},
  title        = {{TOMPEI-CMMD} Dataset},
  howpublished = {The Cancer Imaging Archive},
  version      = {1},
  note         = {[Dataset]},
  publisher    = {The Cancer Imaging Archive},
  year         = {2025},
  doi          = {10.7937/wezw-bh22},
  url          = {https://doi.org/10.7937/wezw-bh22}
}

Citation files: kashiwada2025tompeicmmd.bib (BibTeX) · kashiwada2025tompeicmmd.ris (RIS / EndNote / Zotero)

Summaries are written by the site owner from the published article or dataset record. Refer to the original for authoritative details.