STS-3D-Tooth

The 3D Cone-Beam CT (CBCT) subset of the STS (Semi-supervised Teeth Segmentation) multi-modal dental dataset, as released in Wang et al., Scientific Data 12, 117 (2025) and used in the MICCAI 2023/2024 STS Challenges.

The companion 2D panoramic X-ray subset is hosted at Angelou0516/STS-2D-Tooth.

Dataset Summary

Field	Details
Modality	Cone-Beam CT (CBCT), NIfTI (.nii.gz)
Body Part	Teeth (32 permanent teeth, FDI numbering)
Volumes	371 total: 32 labeled, 339 unlabeled
Volume shape	512 x 512 x 400 (consistent across all volumes)
License	CC-BY-4.0
Source	https://zenodo.org/records/10597292

Subsets

The release ships two distinct CBCT subsets that differ in field-of-view, intensity representation, and label semantics. They are not interchangeable.

Integrity (whole-FOV scan)

Full head/jaw CBCT acquisitions captured in their original field of view.

• 10 labeled volumes (image + binary tooth-vs-background mask)
• 231 unlabeled volumes (image only, no GT)
• Image dtype: float32, intensity range [0.0, 1.0] (pre-normalized to unit interval)
• Affine: identity (1.0 x 1.0 x 1.0 voxel spacing) — true acquisition spacing is not preserved
• Mask labels: {0, 1} — binary foreground = teeth

ROI (tooth-region crop)

Cropped sub-volumes centered on the dental arch.

• 22 labeled volumes (image + multi-class instance mask)
• 108 unlabeled volumes (image only, no GT)
• Image dtype: int16, intensity range approximately [-1000, 3095] (raw HU-like)
• Affine: real isotropic spacing, approximately 0.156 x 0.156 x 0.150 mm
• Mask labels: integer class indices > 0 — per-tooth instance labels following the FDI numbering convention (not all 32 teeth appear in every volume)

Subset	Total	Labeled	Unlabeled
Integrity	241	10	231
ROI	130	22	108
Total	371	32	339

Recommended Ground Truth

Annotation pipeline (per the source paper):

1. 10 junior dentists each annotated CBCT scans layer-by-layer in ITK-SNAP.
2. 3 senior dentists (>10 years experience) reviewed and corrected the layer-wise annotations.
3. Remaining inter-reviewer discrepancies were resolved by consensus.

The shipped masks are post-consensus refined and reflect senior-expert agreement. There is no alternative mask source.

Data Structure

STS-3D-Tooth/
|-- README.md
|-- Integrity/
|   |-- Labeled/
|   |   |-- Image/Integrity_L_NNN.nii.gz   # 10 CBCT volumes (float32, [0,1])
|   |   `-- Mask/Integrity_L_NNN.nii.gz    # 10 binary masks
|   `-- Unlabeled/
|       `-- Image/Integrity_U_NNN.nii.gz   # 231 unlabeled CBCT volumes
`-- ROI/
    |-- Labeled/
    |   |-- Image/ROI_L_NNN.nii.gz         # 22 CBCT crops (int16, raw HU-like)
    |   `-- Mask/ROI_L_NNN.nii.gz          # 22 multi-class FDI instance masks
    `-- Unlabeled/
        `-- Image/ROI_U_NNN.nii.gz         # 108 unlabeled CBCT crops

Image and mask filenames match exactly within each Labeled/ directory.

Splits

The released dataset has no official train/val/test split — define your own downstream. The labeled / unlabeled distinction is intrinsic to the semi-supervised challenge format; the unlabeled volumes have no ground truth and are typically used only for self-training or pretext objectives.

Citation

@article{wang2025sts,
  title   = {A multi-modal dental dataset for semi-supervised deep learning image segmentation},
  author  = {Wang, Yaqi and others},
  journal = {Scientific Data},
  volume  = {12},
  pages   = {117},
  year    = {2025},
  doi     = {10.1038/s41597-024-04306-9}
}

@article{wang2024stschallenge,
  title   = {STS MICCAI 2023 Challenge: Grand challenge on 2D and 3D semi-supervised tooth segmentation},
  author  = {Wang, Yaqi and others},
  journal = {arXiv:2407.13246},
  year    = {2024}
}

License

CC-BY-4.0 (per the Zenodo release at zenodo.org/records/10597292).