tirandazdylan/dentex-panoramic

DENTEX panoramic radiographs: tooth enumeration and diagnosis

Adult dental panoramic radiographs from the labelled training data of the DENTEX challenge (MICCAI 2023), in one schema with film-level splits.

• enumeration: every tooth boxed and numbered (FDI quadrant and position).
• diagnosis: every abnormal tooth boxed and numbered, with its diagnosis: Impacted, Caries, Periapical Lesion or Deep Caries. A tooth with two diagnoses has two rows with nearly the same box. 27 films have no abnormal tooth.

Config	Train	Validation	Test	Boxes (train / validation / test)
enumeration	484	73	66	13,798 / 2,074 / 1,903
diagnosis	565	66	74	2,879 / 328 / 322

Fields

• image: the film as grayscale JPEG.
• image_id: the source file name without .png. The two source tiers number their files independently, so the same image_id in the two configs is usually a different film.
• film_id: the film. Use it to match films across configs.
• width, height: pixels.
• objects.bbox: [x, y, width, height] in pixels, from the top-left corner.
• objects.quadrant: FDI quadrant, 1-4 (1 = the patient's upper right, on the left of the image).
• objects.tooth: position in the quadrant, 1-8 from the midline. The FDI number is 10 * quadrant + tooth.
• objects.diagnosis (diagnosis only): 0 Impacted, 1 Caries, 2 Periapical Lesion, 3 Deep Caries.

Changes from the source

• Source: training_data.zip of DENTEX CHALLENGE 2023, CC BY 4.0. Only the quadrant-enumeration (634 files) and quadrant-enumeration-diagnosis (705 files) tiers are used; the quadrant-only tier, the unlabelled validation films and the unlabelled pre-training films are not.
• PNG films re-encoded as grayscale JPEG, quality 92.
• COCO category_id_1, category_id_2 and category_id_3 renamed to quadrant (+1), tooth (+1) and diagnosis. Boxes unchanged.
• Files grouped into films: two files are one film when their 64x32 thumbnails differ by less than 4 grey levels on average. 260 films are in both configs (the source repeats them under other file names). 11 repeated files within a config are dropped, keeping the first by file name.
• Splits are made per film (80/10/10 by a hash of film_id) and shared by both configs, so no film is in the train split of one config and the test split of the other. The source has no patient IDs, so two different films of one patient can fall in different splits.

Citation

Please cite the DENTEX papers:

@misc{hamamci2023dentex,
  title={DENTEX: Dental Enumeration and Tooth Pathosis Detection Benchmark for Panoramic X-ray},
  author={Hamamci, Ibrahim Ethem and Er, Sezgin and Durugol, Omer Faruk and Cakmak, Gulsade Rabia and de la Rosa, Ezequiel and Simsar, Enis and others},
  year={2023},
  eprint={2305.19112},
  archivePrefix={arXiv}
}
@inproceedings{hamamci2023diffusion,
  title={Diffusion-Based Hierarchical Multi-Label Object Detection to Analyze Panoramic Dental X-rays},
  author={Hamamci, Ibrahim Ethem and Er, Sezgin and Simsar, Enis and Sekuboyina, Anjany and Gundogar, Mustafa and Stadlinger, Bernd and Mehl, Albert and Menze, Bjoern},
  booktitle={Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  year={2023}
}