AU-OPG: Panoramic Dental Radiographs With Oriented Tooth-Level Annotations

AU-OPG (Ajman University Orthopantomography) is a dataset of 901 panoramic dental radiographs annotated for:

• oriented tooth detection;
• radiographic diagnosis; and
• radiographic-evidence-based treatment planning.

The release contains 7,006 tooth-level annotations. Each annotated tooth has a tooth-aligned oriented bounding box, one diagnostic condition, and one corresponding treatment label. The predefined train, validation, and test splits are provided as CSV files.

Clinical scope: AU-OPG supports research on radiographic decision assistance. Its labels describe findings and treatment suggestions consistent with the visible evidence in a panoramic radiograph. They are not complete clinical diagnoses or standalone treatment recommendations.

Links

• Paper: Guiding Neural Networks Toward Better Decisions: A Dental Diagnosis and Treatment Planning Framework and Dataset
• Code and experiments: github.com/akvnn/guiding-neural-nets
• Dataset: huggingface.co/datasets/YSFF/AU-OPG

Annotated example

Dataset composition

Radiographs were randomly collected from the dental clinic at Ajman University after institutional ethics approval. Pediatric, edentulous, and serial radiographs were excluded. Distorted radiographs were retained to better reflect real-world variation. All images were acquired using a Cranex X-ray system.

Split	Radiographs	Tooth annotations
Train	630	4,964
Validation	135	1,010
Test	136	1,032
Total	901	7,006

Diagnostic labels

Condition	Instances
Cavitated	815
Crowned	1,377
Filled	2,857
Impacted	472
Implant	544
Retained Root	941

Treatment labels

Treatment	Instances
None	4,769
Extraction	1,648
Filling	263
Root Canal	326

In the CSV files, an empty value in the treatment column represents the paper's None treatment category.

Annotation process

Annotations were created in Label Studio using a two-stage workflow:

1. dental graduates produced the initial annotations; and
2. an experienced radiologist reviewed every annotation and corrected deviations from the annotation criteria described in the paper.

Bounding boxes were aligned to each tooth's long axis rather than restricted to axis-aligned rectangles. Every annotated tooth has the same oriented box for its condition and treatment labels.

Only teeth belonging to one of the selected diagnostic or treatment categories were annotated. Consequently, unannotated teeth can appear in a radiograph and should not be interpreted as missing annotations for an exhaustive tooth-enumeration task.

Data files

AU-OPG/
├── dataset/                    # Panoramic radiographs (.jpg)
├── train_annotations.csv
├── val_annotations.csv
├── test_annotations.csv
├── sample_with_ground_truth.png
└── README.md

Each CSV row represents one annotated tooth.

Column	Type	Description
image	string	Radiograph filename. The corresponding file is stored at dataset/<image>.
image_width	integer	Original image width in pixels.
image_height	integer	Original image height in pixels.
x	float	X coordinate of the box's top-left corner before rotation, as a percentage of image width.
y	float	Y coordinate of the box's top-left corner before rotation, as a percentage of image height.
width	float	Box width as a percentage of image width.
height	float	Box height as a percentage of image height.
rotation	float	Clockwise rotation in degrees.
condition	string	One of the six diagnostic labels listed above.
treatment	string/null	Extraction, Filling, Root Canal, or empty for None.

The box fields follow Label Studio's percentage-based rectangle representation. Convert the unrotated rectangle dimensions to pixels as follows:

pixel_x = x / 100.0 * image_width
pixel_y = y / 100.0 * image_height
pixel_width = width / 100.0 * image_width
pixel_height = height / 100.0 * image_height

Apply rotation clockwise to recover the tooth-aligned box.

Citation

If you use AU-OPG or the accompanying code, please cite:

@article{alabd_aljabar2026guiding,
  author  = {Alabd-Aljabar, Ahmed and Irshaid, Yousef and Elsayed, Karim and Shaman, Maryam and Arif, Omar and Luke, Alexander M.},
  title   = {Guiding Neural Networks Toward Better Decisions: A Dental Diagnosis and Treatment Planning Framework and Dataset},
  journal = {IEEE Access},
  year    = {2026},
  volume  = {14},
  pages   = {90109--90121},
  doi     = {10.1109/ACCESS.2026.3703711}
}