Preprocessed 2D image dataset derived from the BCS-DBT (Breast Cancer Screening – Digital Breast Tomosynthesis) collection. Contains two resolution variants of 2D images extracted from 3D DBT volumes, with per-image classification labels and bounding-box detection annotations in COCO format.

Dataset Variants

Variant	Folder	Resolution	Image type	Total size
bcs-dbt-2d-low	`bcs-dbt-2d-low/`	~970 × 1229 px	Averaged 2D projection per view	14 GB
bcs-dbt-2d-high	`bcs-dbt-2d-high/`	~1940 × 2457 px	Representative slice per view	50 GB

Both variants cover the same 5,060 patients / 5,610 studies with identical splits and annotations.

Dataset Statistics

Splits

Split	Images
train	19,148
val	1,163
test	1,721
Total	22,032

Classification Labels (per image)

Label	ID	Train	Val	Test
normal	0	18,232	928	1,356
actionable	1	716	160	244
benign	2	124	38	61
cancer	3	76	37	60

Detection Annotations (annotated images only)

Split	Annotated images	Bounding boxes
train	200	224
val	75	75
test	121	136

Detection categories: benign (id=1), cancer (id=2).

Views

Each study contributes up to 4 standard mammography views: lcc, lmlo, rcc, rmlo.

Repository Structure

bcs-dbt-2d-high/
├── images/
│   ├── images_train_part00.tar.gz  …  images_train_part09.tar.gz   (10 shards)
│   ├── images_val_part00.tar.gz
│   └── images_test_part00.tar.gz
├── metadata.zip
│   └── [classification/  detection/  manifest.csv]
├── classification/
│   ├── train.csv
│   ├── val.csv
│   └── test.csv
├── detection/
│   ├── train_coco.json
│   ├── val_coco.json
│   └── test_coco.json
└── manifest.csv

bcs-dbt-2d-low/
├── images/
│   ├── images_train_part00.tar.gz  …  images_train_part02.tar.gz   (3 shards)
│   ├── images_val_part00.tar.gz
│   └── images_test_part00.tar.gz
├── metadata.zip
│   └── [classification/  detection/  manifest.csv  manifest.parquet]
├── classification/
├── detection/
├── manifest.csv
└── manifest.parquet

Each *.tar.gz shard is independently extractable and under 5 GB. metadata.zip bundles all annotation files for convenience.

Usage

Extract all archives

python3 extract_all.py               # extract both variants
python3 extract_all.py bcs-dbt-2d-high   # one variant only
python3 extract_all.py --dry-run     # preview without writing

This extracts images into images/train/, images/val/, images/test/ and metadata into classification/ and detection/.

Manual extraction

# Metadata
unzip bcs-dbt-2d-high/metadata.zip -d bcs-dbt-2d-high/
# strip the "metadata/" prefix manually or use extract_all.py

# Images (each shard independently)
mkdir -p bcs-dbt-2d-high/images/train
tar -xzf bcs-dbt-2d-high/images/images_train_part00.tar.gz -C bcs-dbt-2d-high/images/train/
# repeat for remaining shards

Load classification CSV

import pandas as pd

df = pd.read_csv("bcs-dbt-2d-high/classification/train.csv")
print(df.columns.tolist())
# ['patient_id', 'study_uid', 'view', 'view_base', 'view_type', 'laterality',
#  'split', 'label', 'label_id', 'n_slices', 'slice_idx', 'is_extra_view',
#  'has_annotation', 'flip_applied', 'p_low', 'p_high', 'png_path']

Key columns:

png_path — relative path to the image file (e.g. images/train/DBT-P00013_DBT-S00163_rmlo_s016.png)
label / label_id — classification target (normal=0, actionable=1, benign=2, cancer=3)
p_low, p_high — percentile pixel values for windowing/normalization
has_annotation — whether a bounding box annotation exists in the detection JSON

Load detection annotations (COCO format)

import json

with open("bcs-dbt-2d-high/detection/train_coco.json") as f:
    coco = json.load(f)

# coco["categories"]: [{"id": 1, "name": "benign"}, {"id": 2, "name": "cancer"}]
# coco["annotations"][0]:
#   {"id": 1, "image_id": 1, "category_id": 1,
#    "bbox": [x, y, w, h], "area": ..., "iscrowd": 0}

Image loading

Images are 16-bit grayscale PNG. Use p_low / p_high from the CSV for windowing:

import numpy as np
from PIL import Image

img = np.array(Image.open("bcs-dbt-2d-high/images/train/DBT-P00013_DBT-S00163_rmlo_s016.png"))
# img.dtype == uint16, shape == (H, W)

row = df[df["png_path"] == "images/train/DBT-P00013_DBT-S00163_rmlo_s016.png"].iloc[0]
img_clipped = np.clip(img, row["p_low"], row["p_high"])
img_norm = (img_clipped - row["p_low"]) / (row["p_high"] - row["p_low"])

Image Filename Convention

DBT-P{patient_id}_DBT-S{study_uid}_{view}_{suffix}.png

bcs-dbt-2d-high:  suffix = s{slice_idx:03d}   e.g. s016
bcs-dbt-2d-low:   suffix = avg                 (averaged projection)

Source Dataset

Original collection: BCS-DBT on TCIA
Reference paper: Buda et al., Data from Breast Cancer Screening – Digital Breast Tomosynthesis (BCS-DBT), 2021. DOI: 10.7937/e4wt-cd02
License: Academic Free License 3.0

Citation

If you use this dataset, please cite the original BCS-DBT collection:

@misc{buda2021bcsdbt,
  author    = {Buda, Mateusz and others},
  title     = {Data from Breast Cancer Screening – Digital Breast Tomosynthesis (BCS-DBT)},
  year      = {2021},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/e4wt-cd02}
}

Downloads last month: 94

Total file size:

67 GB