ToothMaker: Realistic Panoramic Dental Radiograph Generation via Disentangled Control.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• High-fidelity dental radiographs are crucial for training diagnostic AI models.
• Existing generative methods struggle with dental radiology's complexity, leading to inaccurate structures and concepts.
• Generative approaches for dental radiology remain largely unexplored.

Purpose of the Study:

• To investigate diffusion-based teeth X-ray image generation for the dental domain.
• To introduce ToothMaker, a novel framework for high-fidelity dental image synthesis.
• To improve the accuracy and realism of generated dental radiographs.

Main Methods:

• Developed ToothMaker, a framework employing diffusion-based generation.
• Introduced control-disentangled fine-tuning (CDFT) for style and layout control.
• Proposed prior-disentangled guidance module (PDGM) using LLMs and hypergraph networks for dental concept synthesis.

Main Results:

• ToothMaker successfully synthesizes high-fidelity and diverse dental X-ray images.
• The generated data significantly improved performance in downstream segmentation and visual question answering tasks.
• The approach demonstrated a reduced reliance on manually annotated data.

Conclusions:

• ToothMaker is the first diffusion-based framework specifically designed for dental X-ray image generation.
• The proposed CDFT and PDGM strategies effectively address challenges in dental image synthesis.
• Generated data from ToothMaker can enhance diagnostic model training and reduce annotation efforts.