OralSeg

Model Card for OralSeg

<!-- Provide a quick summary of what the model is/does. -->

OralSeg focuses on tooth and bone object detection and segmentation in dental CBCT imaging, and is applicable to surgical planning, orthodontic design, implant planning, and academic research.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

OralSeg is an advanced instance segmentation model based on dental CBCT, developed in our study, designed to accurately segment the maxilla, mandible, 32 teeth, and bilateral mandibular canals in large-scale dental CBCT images. The model is trained using high-precision annotations from expert dental professionals and is characterized by its efficiency, robustness, and accuracy, making it well-suited for clinical research and practical applications.

• Developed by: AIADIR
• Funded by: The University of Hong Kong, Faculty of Dentistry
• Model type: Dental CBCT image instance segmentation model
• Language: English
• License: OralSeg © 2025 by AIADIR is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). The model is released for non-commercial use.

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

OralSeg was trained and tested using 100 high-quality, high-resolution dental CBCT images. All images were manually annotated at pixel-level precision by expert dental professionals, covering:

• Maxilla
• Mandible
• 32 teeth (including wisdom teeth)
• Left and right mandibular canals

These detailed annotations provide the network with accurate and rich target contour information during training, significantly improving the model's segmentation accuracy and robustness.

Model Architecture

OralSeg model is based on the SwinUnetR architecture and incorporates our proprietary data augmentation strategies and training schemes to further enhance its performance in dental instance segmentation tasks. This architecture features the following key components:

• Multi-scale feature extraction: Uses a Swin Transformer-based feature extraction module to process anatomical structures of various sizes in CBCT images.
• UNet-style convolutional structure: Features extracted by the encoder are progressively upsampled and fused via a symmetric decoder path, preserving spatial information and improving segmentation accuracy.
• Transformer self-attention mechanism:: Transformer self-attention mechanism: Effectively enhances the ability to model long-range pixel dependencies, improving detail capture and complex structure segmentation in high-resolution medical images.

Application Scenarios

• Clinical research: Assists clinicians in quickly identifying and segmenting key structures such as teeth, bones, and nerve canals during CBCT examinations and surgical planning.
• Implant and orthodontic planning: Enables more convenient and visualized treatment planning, preoperative assessment, and simulation.
• Digital education and training: Provides standardized 3D anatomical segmentation results for medical education and research.

Disclaimer and Compliance

• Licensing: This model is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Users are free to copy, distribute, display, and adapt the model for non-commercial purposes, including research, education, and medical training. When using this model, users must credit the original developers (AIADIR, Faculty of Dentistry, The University of Hong Kong).
• Data compliance: All training data used by the model were annotated by professional dental experts, sourced legally and ethically, with patient privacy fully protected.
• Potential bias: While the model performs well across various types and structures of teeth and bones, there may still be rare cases or distributional biases not fully covered. Further evaluation or fine-tuning is recommended for specific scenarios.
• Safety notice: Model outputs are for reference only and should not replace professional clinical diagnosis or decision-making. All use cases should be accompanied by expert dental evaluation and judgment.