Spatial Prior-Guided Bi-Directional Cross-Attention Transformers for Tooth Instance Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate tooth instance segmentation in dental panoramic X-rays is clinically vital.
• Existing methods often neglect spatial prior information, leading to misidentification of similar teeth.
• This limitation impacts dental diagnosis, treatment planning, and research.

Purpose of the Study:

• To propose SPGTNet, a spatial prior-guided transformer network for accurate tooth instance segmentation.
• To leverage both CNN-extracted positional features and vision transformer contextual information.
• To improve the identification of tooth categories, especially for adjacent or similarly shaped teeth.

Main Methods:

• SPGTNet integrates CNNs and vision transformers with a novel spatial prior perception module.
• A center-based module enhances spatial prior information for CNN features.
• A bi-directional cross-attention module fuses CNN and transformer features.

Main Results:

• SPGTNet demonstrated superior performance on three public benchmark datasets.
• The method accurately identifies and analyzes tooth structures.
• It outperforms existing state-of-the-art approaches in tooth instance segmentation.

Conclusions:

• SPGTNet effectively utilizes spatial prior information for improved tooth segmentation.
• The proposed method offers a significant advancement for dental diagnostic tools.
• Accurate segmentation provides crucial data for clinical decision-making and research.