metadata

license: cc0-1.0
task_categories:
  - image-classification
  - image-segmentation
tags:
  - medical
pretty_name: T-SYNTH
size_categories:
  - 1K<n<10K

T-SYNTH

T-SYNTH is a synthetic digital breast tomosynthesis (DBT) dataset with four breast fibroglandular density distributions imaged using Monte Carlo x-ray simulations with the publicly available Virtual Imaging Clinical Trial for Regulatory Evaluation (VICTRE) toolkit.

Dataset Details

The dataset has the following characteristics:

Breast density: dense, heterogeneously dense, scattered, fatty
Mass radius (mm): 5.00, 7.00, 9.00
Mass density: 1.0, 1.06, 1.1 (ratio of mass radiodensity to that of fibroglandular tissue)

Dataset Description

Curated by: Christopher Wiedeman, Anastasiia Sarmakeeva, Elena Sizikova, Daniil Filienko, Miguel Lago, Jana Gut Delfino, Aldo Badano
License: Creative Commons 1.0 Universal License (CC0)

Data Acquisition Details

Imaging Modality: Paired 2D digital mammography (DM) and 3D digital breast tomosynthesis (DBT) images. The DBT images are projected into C-VIEW via the method of (Klein, 2023).

Manufacturer and Model: Replica of the Siemens detector based on MC-GPU (Badal and Badano, 2009).

Demographics: All breast phantoms are synthetic and do not represent real human subjects.

Cohort Description: 9,000 synthetic digital breast tomosynthesis (DBT) samples, distributed across four breast density categories:

Breast Density	Fatty	Scattered	Hetero	Dense	Total
Les.-free / Les.-present	1350/1350	1350/1350	900/900	900/900	4500/4500

Annotation Protocols: Lesion masks and bounding boxes were generated automatically from the phantom.

Data Format and Structure: Image files are in .raw format.

Dataset Sources

Code: https://github.com/DIDSR/tsynth-release
Arxiv: https://arxiv.org/abs/2507.04038

Intended Use

T-SYNTH is intended to facilitate testing of AI with pre-computed synthetic digital breast tomosynthesis (DBT) data, complementing the M-SYNTH synthetic mammography dataset.

Ethical Considerations

This work is using synthetically generated data and does not include any real patient-identifiable information. Publication of synthetic data aims to promote transparency, reproducibility, and fairness in medical AI research. We recommend avoiding training models only on synthetic data without appropriate validation.

Dataset Structure

Directory layout:

T-SYNTH/data/
├── cview
├── embed_metadata
├── pretrained_models
├── results
└── volumes_subset

Descriptions:

cview/ -- contains T-SYNTH C-VIEW images.
embed_metadata/ -- Configuration files needed to reproduce experiments.
pretrained_models/ -- Pretrained models for DBT, DM and diffusion experiments to reproduce results from the paper. Note to reproduce you need files from embed_metadata/.
results/ -- Output data and plots used in the paper (see T-SYNTH repository). Description for each experiment could be found here.
volumes_subset/ -- example of volumetric data. The full data set will be released later due to volume.

The data is organized by lesion size, breast density and lesion density. Folder names follow the convention: output_cview_det_Victre/device_data_VICTREPhantoms_spic_[LESION_DENSITY]/[BREAST_DENSITY]/2/[LESION_SIZE]/SIM.zip.

Example path in volumes_subset:

device_data_VICTREPhantoms_spic_1.1/fatty/2/5.0/SIM/D2_5.0_fatty.1/1/
├── reconstruction1.loc        # Lesion coordinates
├── reconstruction1.mhd        # Metadata (raw format)
├── reconstruction1.raw        # Raw image data
└── reconstruction1_mask.h5    # Pixel-level segmentation masks for lesions and tissues

How to use it

The description how to use it could be found here.

Citation

@article{t-synth,
  title={T-SYNTH: A Knowledge-Based Dataset of Synthetic Breast Images},
  author={Christopher Wiedeman, Anastasiia Sarmakeeva, Elena Sizikova, Daniil Filienko, Miguel Lago, Jana G. Delfino, Aldo Badano},
  journal={},
  volume={},
  pages={},
  year={2025}
}

Datasets:

didsr
/

tsynth