dmolino's picture
Add `text-to-3d` task category (#1)
33560ac verified
metadata
language:
  - en
license: apache-2.0
size_categories:
  - 1K<n<10K
pretty_name: CT-RATE_Synthetic
tags:
  - medical
task_categories:
  - text-to-3d

Dataset Card for Synthetic Text-to-CT Scans - VLM3D Challenge

Dataset Details

Dataset Description

This dataset contains 1,000 synthetic 3D chest CT scans generated using the model introduced in
Text-to-CT Generation via 3D Latent Diffusion Model with Contrastive Vision-Language Pretraining (Molino et al., 2025).

The model was trained on the CT-RATE dataset, the largest publicly available collection of paired CT volumes and radiology reports.
It leverages a 3D latent diffusion framework combined with contrastive vision-language pretraining (3D-CLIP) to synthesize anatomically coherent and semantically faithful CT scans directly from clinical text prompts.

These 1,000 scans were generated for the VLM3D Challenge - Task 4, serving as a benchmark resource for multimodal evaluation and synthetic data research in medical imaging.

  • Curated by: ArCo Lab – Università Campus Bio-Medico di Roma & Umeå University
  • Language(s): Conditioning report are in English
  • License: Apache 2.0

Dataset Sources

Uses

Direct Use

  • Benchmarking text-to-CT generative models.
  • Data augmentation for classification, detection, or segmentation tasks.
  • Research in multimodal vision-language learning for 3D medical imaging.
  • Educational purposes and simulation in medical training.

Out-of-Scope Use

  • Direct diagnostic or clinical use.
  • Deployment in healthcare without proper validation and regulatory approval.
  • Any attempt to re-identify patients (note: scans are fully synthetic).

Dataset Structure

  • Format: Volumetric CT scans stored in NIfTI (.nii.gz) format.
  • Resolution: Resampled to 0.75 × 0.75 × 3.0 mm voxel spacing, cropped/padded to 512 × 512 × 128.
  • Intensity: Normalized in Hounsfield Units (clipped to [−1000, +1000]).
  • Content: Synthetic chest CT scans across 18 pathological conditions (e.g., nodules, opacities, effusion, emphysema).

Dataset Creation

Curation Rationale

Created to provide a reproducible benchmark for text-to-CT generation and to supply synthetic volumetric data for research in data augmentation, privacy preservation, and multimodal foundation models.

Source Data

  • Trained on CT-RATE (Hamamci et al., 2024), a large-scale dataset of chest CTs paired with radiology reports.

Annotations

No manual annotations included; diagnostic semantics are embedded via the conditioning text prompts used during generation.

Personal and Sensitive Information

  • The dataset contains no real patient data.
  • All scans are synthetic and generated by a model trained on anonymized public datasets.

Bias, Risks, and Limitations

  • Synthetic data may not fully capture rare pathologies or distributional nuances of real-world scans.
  • While useful for augmentation and benchmarking, these scans are not clinically validated.
  • There is a potential risk if synthetic data are used without acknowledging their limitations in medical research.

Recommendations

Users should:

  • Combine synthetic with real-world data for downstream tasks.
  • Avoid over-relying on synthetic volumes for clinical translation.
  • Report the provenance of synthetic data when used in publications.

Citation

If you use this dataset, please cite the following work:

BibTeX:

@article{molino2025textct,
  title={Text-to-CT Generation via 3D Latent Diffusion Model with Contrastive Vision-Language Pretraining},
  author={Molino, Daniele and Caruso, Camillo Maria and Ruffini, Filippo and Soda, Paolo and Guarrasi, Valerio},
  journal={arXiv preprint arXiv:2506.00633},
  year={2025}
}