gulluk/mammosightr-preprocessed

MammosighTR — Preprocessed Mammography Dataset (BI-RADS)

Preprocessed PNG mammograms with image-level BI-RADS labels, derived from the nationwide Turkish breast-cancer screening dataset (MammosighTR) released for the TEKNOFEST 2023 Artificial Intelligence in Health Competition by the Republic of Turkey Ministry of Health. Original DICOMs are cropped to the breast region with a YOLOX detector and exported as PNG; we add an image-level metadata mapping built from the official patient-level annotations.

This dataset accompanies the paper MAM-CLIP: Vision–Language Pretraining on Mammography Atlases for BI-RADS Classification (arXiv:2605.19359). Code: github.com/igulluk/MAM-CLIP.

Usage

from datasets import load_dataset

ds = load_dataset("gulluk/mammosightr-preprocessed", split="train")
ex = ds[0]
ex["image"]          # PIL.Image (cropped mammogram)
ex["birads"]         # 0, 1, 2, 4, or 5
ex["laterality"]     # "left" / "right"
ex["view"]           # "CC" / "MLO"

Fields

Field	Description
image	Cropped mammogram (PNG)
patient_id	Anonymized patient id
image_filename	LCC.png / LMLO.png / RCC.png / RMLO.png
view	CC or MLO
laterality	left or right
birads	Image-level BI-RADS (0, 1, 2, 4, 5)
birads_label	e.g. "BI-RADS 4"
breast_composition	ACR density A/B/C/D
findings_locations	Finding quadrants (English)
patient_birads	Original patient-level BI-RADS

The cohort contains no BI-RADS 3 or BI-RADS 6 cases. 42,074 images from 10,740 patients (840 patients have 3 of the 4 views).

Image-level BI-RADS distribution

BI-RADS	0	1	2	4	5
Count	5,300	18,448	10,088	3,799	4,439

Label construction (patient → image)

Patient-level BI-RADS is mapped from the source "Kategori N" field. SAĞ = right (RCC, RMLO), SOL = left (LCC, LMLO). A quadrant cell is empty if NaN, "" or [].

• Patient BI-RADS == 1 → both breasts BI-RADS 1.
• Patient BI-RADS != 1: findings on both sides → both breasts get the patient score; findings on one side → that side gets the score, the other gets BI-RADS 1; no findings on either side → patient score on both breasts.
• Both views (CC, MLO) of a breast inherit that breast's label.

License

CC BY-NC 4.0 — non-commercial research use with attribution.

Citation — please cite BOTH

This work:

@misc{gulluk2026mamclip,
  title         = {MAM-CLIP: Vision--Language Pretraining on Mammography Atlases for BI-RADS Classification},
  author        = {Halil Ibrahim Gulluk and Olivier Gevaert},
  year          = {2026},
  eprint        = {2605.19359},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.19359}
}

Original source (MammosighTR):

@article{koc2025mammosightr,
  title   = {MammosighTR: Nationwide Breast Cancer Screening Mammogram Dataset with BI-RADS Annotations for Artificial Intelligence Applications},
  author  = {Ko\c{c}, Ural and others},
  journal = {Radiology: Artificial Intelligence},
  volume  = {7}, number = {6}, pages = {e240841}, year = {2025},
  doi     = {10.1148/ryai.240841}, note = {PMID: 40801802}
}

Disclaimer

Research use only. Labels are screening BI-RADS assessments, not biopsy-confirmed outcomes; BI-RADS 0 indicates an incomplete assessment. Quadrant localization is patient-level and propagated by the rule above; it is not a per-image bounding box.