Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This sample is shared for evaluation only. By requesting access, you agree to the terms in TERMS.md. Redistribution, production or published model training, and any attempt to re identify a patient are prohibited. Requests are approved automatically.
Log in or Sign Up to review the conditions and access this dataset content.
MouthCare intraoral dataset
18,000 intraoral photographs from 3,000 patient series, with six clinical classes and dentist authored pixel level annotations.
MouthCare is a clinical image dataset built from a standard six view intraoral protocol. Each photograph is annotated with closed polygons for tooth, gum, tartar, cavity, gingivitis and restoration.
This page provides a gated evaluation sample of 480 images from 80 patients. It is intended for teams that want to inspect the annotation format, image quality and technical fit before discussing access to the full collection.
Access is gated and governed by the Evaluation and Privacy Terms. More information is available at mouthcare.fr/en/dataset.
Evaluation sample at a glance
| Item | Value |
|---|---|
| Intraoral photographs | 480 |
| Patient series | 80 |
| Standardised views per patient | 6 |
| Clinical classes | 6 |
| Pixel level polygon annotations | 3,632 |
The sample includes clean photographs, visual overlays, per image JSON records, multi channel masks, a COCO export, patient level metadata, split definitions and a checksum file.
What the full dataset contains
The complete collection contains 18,000 photographs, six views per patient series, and 122,877 polygons. It was collected in dental practices in India under a single capture protocol. A further 6,000 photographs are currently being collected using the same process.
| Class | Polygons | Images containing the class | Patients with the class |
|---|---|---|---|
| Tooth | 37,225 | 100% | 100% |
| Gum | 50,732 | 98.4% | 100% |
| Tartar | 24,144 | 33.4% | 54.2% |
| Cavity | 7,916 | 25.0% | 49.5% |
| Restoration | 1,711 | 6.3% | 12.0% |
| Gingivitis | 1,149 | 4.5% | 13.9% |
The proportion of images containing a class is generally more useful than its share of polygons. A single image can contain many tooth polygons but only one gum polygon. Across the full dataset, there are 6.8 polygons per image on average.
Evaluation sample
The release contains 480 images from 80 patient series. It includes healthy anatomy, heavier clinical findings and less frequent classes. It is for evaluation and is not a training release.
example.py shows how to load an image, its masks and its annotation record.
The files share the same image identifier. For example, 4_P2148 refers to the same image in images, masks, annotations and overlays.
Sample class distribution
| Class | Polygons | Images | Patient series |
|---|---|---|---|
| Tooth | 975 | 478 | 80 |
| Gum | 1,428 | 471 | 80 |
| Tartar | 548 | 143 | 40 |
| Cavity | 188 | 106 | 36 |
| Gingivitis | 215 | 91 | 24 |
| Restoration | 278 | 119 | 24 |
Interactive overlay viewer
The Dataset Viewer contains 36 exact polygon overlays from six anonymised patient series. Each series contributes its six standard views. The table can be filtered by patient, view, arch, sextant and visible classes.
The viewer includes P6, P804, P2310, P2356, P2364 and P2389. Together, these cases include every released class, including gingivitis and restoration. The remaining images stay within the gated sample.
Annotation method
All polygons in this release were drawn manually by Romain Lateur, a practising dentist and the founder of MouthCare. Each class has a defined clinical extent: tooth and gum follow the visible anatomical boundary, findings are traced to their own edge and drawn on the tissue that carries them, and the six view capture protocol fixes the anatomical position of every annotation.
Quality assurance
Dr Rayane Amine Khodja reviewed the completed annotations to validate the work. Each image also receives pixel based quality checks for blur, exposure and glare, and the delivered metadata records the result for that image.
This is not a multi reader consensus dataset, so no formal inter annotator agreement score is available for this evaluation sample.
Findings are annotated on their supporting tissue. A cavity remains inside a tooth polygon and gingivitis remains inside a gum polygon. The masks therefore hold one independent channel per class. If a single label map is required, resolve overlaps in this order: cavity, gingivitis, tartar, restoration, tooth, then gum.
tooth polygons represent a sextant, or a group of adjacent teeth. Individual tooth numbering is not part of this release.
Image views and sample split
The six views are upper centre, upper right, upper left, lower centre, lower right and lower left. Right and left follow the patient's clinical orientation. In photographs taken while facing the patient, the visual position is therefore reversed.
This split applies only to this 480 image evaluation sample. It does not describe the split of the full 18,000 image dataset. The split is patient disjoint: 70% for training, 15% for validation and 15% for testing. All six views from a patient remain in the same split.
Collection and image quality
Photographs were captured by dentists at five clinical sites in India using iPhones, with natural light or flash. Resolution varies and is recorded for every image. The median resolution across the full dataset is 1.4 MP, and 11.6% of images are 8 MP or above.
The sample flags blur, exposure, glare and sparse framing. Sixty one images are marked sparse_framing. The provenance record states whether an image was resized before delivery. Duplicate checks were performed at patient level to keep the splits separate.
Privacy, consent and rights
Every delivered image has had its EXIF metadata removed. Images are restricted to the oral region and areas outside it are masked. Where masking could not reliably remove a feature, the relevant area was blacked out. The delivery uses pseudonymous identifiers only and contains no patient names, dates of birth, contact details or demographic attributes.
Patients gave signed consent at the time of photography for research, artificial intelligence development and commercial use. The collecting dentists assigned rights to NT Corporation by signed deed. The annotations, masks and classifications were created by NT Corporation and remain its property. Original consent documents are retained by the collecting clinics and are not included in the release.
The dataset includes patients of all ages, including minors. Do not attempt to identify any patient or combine these images with other information for that purpose.
Limitations
Resolution varies, as is common with clinical smartphone photography. Some images have sparse framing or clinical quality flags and should be filtered according to the use case. Annotation is performed by one dentist, so this version does not include a formal inter annotator agreement measure. The sample is useful for evaluation but is too small to represent the complete clinical distribution or to support training.
Access and citation
The sample is available for evaluation only. Redistribution, model training for production or published research, and re identification attempts are prohibited. Feel free to request access to review the sample. A separate licence is required for production, research or redistribution use. Contact contact@mouthcare.fr.
NT Corporation, MouthCare. MouthCare Intraoral Dataset, evaluation sample v2.2 (2026).
https://www.mouthcare.fr/en/dataset
- Downloads last month
- 59