ToothFairy3
CBCT dataset of 532 NIfTI volumes with comprehensive dental annotations covering 77 classes, released for the ToothFairy3 challenge at ODIN2025.
Overview
This is the dataset associated with the ToothFairy3 challenge, organized by the University of Modena and Reggio Emilia in collaboration with Radboud University. The challenge was hosted by grand-challenge and is part of the ODIN2025 workshop within MICCAI2025. For challenge details, please visit the dedicated website. The dataset paper has been published at MICCAI2026.
Features
Although carrying additional annotations, the volumes of Set A (leading P in the file name) and Set B (leading F) overlap with those of the ToothFairy2 dataset. Data from Set B has a broader field of view, including the complete upper-teeth segmentations, and the acquisition machine for Set A and Set B is the same. Data from Set C (leading S) were not previously released with the ToothFairy2 dataset and were acquired with a different machine; the field of view of Set C is similar to that of Set A.
What’s new in ToothFairy3:
- 🦷 52 new CBCT volumes, acquired using a different scanner;
- 🧠 35 additional labels, including pulpy cavities for all 32 teeth, as well as left/right incisive canals and the lingual canal;
- ✅ Improved annotations from ToothFairy2.
| Field | Value |
| # Set A | 417 |
| # Set B | 63 |
| # Set C | 52 |
| # Classes | 77 |
| Format | NIfTI |
| Scale | Hounsfield |
Dataset Structure
The ToothFairy3 dataset follows the nnU-Net dataset format. It consists of three components: raw images, the corresponding segmentation maps, and a dataset.json file specifying some metadata. Data are available in NIfTI format.
The class IDs are an extension of the FDI World Dental Federation notation and include background, jaws, inferior alveolar canals, maxillary sinus, pharynx, bridges, crowns, implants, upper and lower teeth (wisdom teeth included), incisive canals, the lingual canal, and teeth roots/pulps, for a total of 77 classes.
Radiological Orientation
This dataset uses the RPI orientation, which is not a standard convention, but is consistent with the first release of the Maxillo dataset.
If you wish to convert the dataset to the standard LPS orientation, you can use this Python script. The script requires two input parameters:
- the path to the dataset (with
imagesandlabelssubfolders), and - the output path where the transformed files will be saved.
Note: if you are participating in the ToothFairy challenge, this transformation is not required. The orientation of all test data matches that of the training data as provided on this website.