FakeBraiNCCT

Fully synthetic brain non-contrast CT (NCCT) slices of subacute ischemic stroke, with their multi-class segmentation masks. The dataset contains 10,000 synthetic masks, each paired with one SPADE image and one SegDiff image (20,000 synthetic images in total). No real patient image is included.

It accompanies the manuscript "A Multi-Dimensional Comparison of Mask-Conditioned Generators for Brain NCCT Synthesis in Subacute Ischemic Stroke" (Borrego et al., under review). Evaluation code: https://github.com/lucixia3/synthetic_brain_evaluation

Dataset structure

Field	Type	Description
id	string	Sample identifier
mask	image	Synthetic 7-class segmentation mask (mutually exclusive classes)
spade	image	NCCT slice generated by SPADE from mask
segdiff	image	NCCT slice generated by SegDiff from mask (background-corrected)
label	int	Lesion presence: 1 = ischemic infarct present in the mask, 0 = no infarct

Single split (train), 10,000 rows: 5,000 lesion and 5,000 non-lesion masks.

Mask classes: 0 background, 1 soft tissue, 2 bone, 3 CSF, 4 white matter, 5 grey matter, 6 ischemic infarct.

Generation procedure

1. Masks. Generated by an anatomy-aware latent diffusion model (MaskVAE + conditional LDM) conditioned on the binary lesion label (arXiv:2602.10167). The mask generator was trained only on masks of the 61 training patients.
2. Images. Each mask was given to two conditional generators trained on the same 61 patients:
   • SPADE (GAN with spatially-adaptive normalization).
   • SegDiff (segmentation-guided DDPM, DDIM sampling with 50 steps). SegDiff outputs had a systematic background offset; pixels labelled background in the mask were set to 0, as in the real images.

Preprocessing (source data and outputs)

• Source: follow-up NCCT acquired 24–48 h after onset in patients with cardioembolic ischemic stroke treated with mechanical thrombectomy (single centre, Siemens scanners, slice thickness < 2 mm, standard brain kernel).
• Brain window (width 80 HU, level 40 HU), 8-bit.
• Supratentorial axial slices (20–95% of cranial height).
• Images and masks at 256 × 256 pixels.

Provenance of the conditioning masks

The masks are synthetic, but the mask generator learned from real masks: CTseg automated tissue segmentation, quality-controlled by a neuroradiologist, combined with expert manual infarct delineation, from 61 training patients. The study was approved by the Ethics Committee of the Institut de Recerca Sant Pau (IIBSP-IIC-2025-108), with a waiver of informed consent.

Privacy assessment

Images generated from the synthetic masks were audited with an embedding-based near-duplicate detector (DINOv2 ViT-S/14 with a triplet-trained projection head) against all 5,238 training slices. The threshold was the 5th percentile of validation-to-training distances. Flagged rates: SegDiff 1.2%, SPADE 0.1%, both below the 3.8% rate of unseen real test patients. Flagged images share coarse head shape with their nearest training slice but differ in parenchymal detail.

This audit shows no evidence of near-duplicate reproduction under this detector. It does not exclude other forms of information leakage (e.g. membership inference) and is not a legal assessment of data protection compliance.

Intended uses

• Research on synthetic data for medical image analysis, e.g. pre-training or data augmentation for brain tissue segmentation.
• Benchmarking image-quality, radiomic and memorization metrics for generative models.
• Teaching.

Out-of-scope uses

• Clinical diagnosis or any decision about real patients.
• Training infarct-detection models intended for clinical use: in our experiments, infarct segmentation was not learnable from these 2D slices.
• Commercial use (licence is non-commercial).

Limitations

• Single centre, single stroke subtype (cardioembolic), post-thrombectomy follow-up scans, 24–48 h window.
• 2D axial slices, supratentorial range only, pre-windowed 8-bit intensities (HU values cannot be recovered).
• Hemorrhagic transformation and contrast staining present in the source data were not modelled as separate classes.
• SegDiff reproduces imperfections of the synthetic masks; SPADE images are smoother and less diverse.
• Synthetic images did not improve tissue segmentation when added to a large real training set; they were most useful when synthetic data were the main training source.

Licence

CC BY-NC 4.0

Funding

Supported by the Innovative Health Initiative Joint Undertaking (IHI JU) under grant agreement No. 101172997 (SEARCH).

Citation

If you use this dataset, please cite the accompanying manuscript (citation will be updated on publication) and the mask-generation preprint:

@article{borrego2026anatomy,
  title   = {Anatomy-preserving latent diffusion for generation of brain segmentation masks with ischemic infarct},
  author  = {Borrego, Lucia and Thambawita, Vajira and Ciuffreda, Marco and del Val, Ines and Dominguez, Alejandro and Munuera, Josep},
  journal = {arXiv preprint arXiv:2602.10167},
  year    = {2026}
}