Towards Better Cephalometric Landmark Detection With Diffusion Data Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthodontics

Background:

• Cephalometric landmark detection is crucial for orthodontic diagnostics and treatment planning.
• Limited sample sizes and manual annotation efforts hinder the creation of diverse datasets for deep learning models.
• This scarcity restricts the effectiveness of advanced deep learning methods, especially large-scale vision models.

Purpose of the Study:

• To develop an automated data generation method for diverse cephalometric X-ray images and annotations.
• To address the limitations of data scarcity and manual annotation in cephalometric analysis.
• To enhance the accuracy of deep learning-based cephalometric landmark detection.

Main Methods:

• An innovative approach using anatomical priors to construct cephalometric landmark annotations.
• A diffusion-based generator to create realistic X-ray images aligned with annotations.
• Introduction of a novel prompt cephalometric X-ray image dataset for controlled attribute generation.
• Integration of large-scale vision detection models with generated data for improved accuracy.

Main Results:

• The proposed method successfully generates diverse cephalometric X-ray images and annotations without human intervention.
• Training with generated data significantly enhances the performance of cephalometric landmark detection models.
• The Success Detection Rate (SDR) improved by 6.5%, reaching 82.2% compared to baseline methods.

Conclusions:

• The automated data generation method effectively overcomes data limitations in cephalometric analysis.
• The approach enables the application of large-scale vision models for improved orthodontic diagnostic accuracy.
• This work provides a valuable resource for advancing AI in dental and orthodontic applications.