TOMPEI-CMMD

Pixel-level breast-lesion segmentation on full-field digital mammography (FFDM). 1385 annotated breasts / 1363 patients, MLO view only.

What this mirror is

Upstream TOMPEI-CMMD ships annotations only — 1385 polygon JSON files plus a clinical spreadsheet. The images belong to the separate CMMD collection on TCIA. This repository publishes the joined product: each row pairs a CMMD mammogram with the TOMPEI polygon masks rasterised onto it.

The join key is (PatientID, ImageLaterality (0020,0062), ViewCodeSequence (0054,0220) == 399368009 "medio-lateral oblique"). Never file order — instance order varies between series, and ViewPosition (0018,5101) is absent from these DICOMs.

Mask semantics — read before benchmarking

• mask_lesion — union of every polygon, built directly from the polygon list. It is not composited from the per-class columns, so no paint-order bias exists.
• mask_<class> — null means CONFIRMED NEGATIVE, not "unannotated." Annotation is complete for every Malignant/Benign breast, so a missing class is a positive statement that the class is absent. This is the inverse of datasets such as DRAC22, where an absent mask means the class was never labelled. Per-class targets therefore score only the positive breasts under point-prompt evaluation, while the confirmed negatives remain in the mirror.

Classes overlap — 276 breasts carry ≥2 distinct classes (calc+mass alone = 238), so they are published as separate binary masks rather than one multi-class map, which would erase pixels.

class	breasts	role
mass	1013	headline target
calc	541	headline target
dist	60	reportable
FAD	33	reportable, not a headline number
FA	22	reportable, not a headline number
lipoma	1	published, excluded from the default target set

Known data characteristics

• 4 CMMD laterality-tag swaps. For D1-0999_R, D2-0224_R, D2-0229_R and D2-0642_L the CMMD DICOM ImageLaterality contradicts the image content. TOMPEI's side label is the anatomically correct one, so the annotation is paired with the instance whose content matches. Flagged by laterality_tag_swapped; the original tag is preserved in dicom_tag_laterality. These were found by pixel evidence, not by any upstream flag — the clinical sheet does not record them.
• dist extends past the parenchymal edge. Architectural distortion is outlined across the retracted zone, so 7 dist breasts have frac_in_tissue < 0.90 (min 0.598). Every non-dist breast is ≥ 0.984. This is annotation semantics, not mis-pairing. frac_in_tissue is recorded per sample so it can be filtered.
• 3 breasts are classification == "Exclusion" (D1-0985_L, D1-1691_R, D2-0367_L) yet still carry annotations. They are retained here and flagged with is_exclusion; filter them out to reproduce the paper's 1382-breast figure.
• Polygon vertices run out of bounds (x up to 1923 against width 1914) and are clipped.
• The upstream label field has whitespace variants ('calc ' vs 'calc'); stripped here. The type field is the drawing primitive, not the class, and color is unusable as a class key (1144 distinct hex values for 6 classes).

Overlap and leakage

100 % patient overlap with CMMD by construction — the images are CMMD. The cross-reference is patient_id (D1-#### = CMMD1 benign/malignant subset, D2-#### = CMMD2 molecular-subtype subset). Anything benchmarking on CMMD leaks against this set. No overlap with CBIS-DDSM, INbreast, VinDr-Mammo, MIAS or CDD-CESM.

Splits

No official train/val/test split exists upstream; everything is published as train. Group on patient_id when splitting — 22 patients contribute both breasts.

Licence

CC BY 4.0. The annotations are CC BY 4.0, and the CMMD images are independently CC BY 4.0 — verified across all 1775 CMMD series via the TCIA API, which reports Creative Commons Attribution 4.0 International License for every one. The joined product takes the more restrictive of the two licences, which here is still CC BY 4.0, so there is no non-commercial downgrade.

Citation

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

Images: Cui, C. et al. The Chinese Mammography Database (CMMD) [Dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/tcia.eqde-4b16

Note: a 2026 Neurocomputing article titled "TOMPEI-CMMD: An adaptive radiological feature analysis using unitary U-net…" is by unrelated authors reusing the name. It is a method paper, not the data descriptor. Cite the TCIA DOI.