Two-Stream Graph Convolutional Network for Intra-Oral Scanner Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Geometry

Background:

• Accurate tooth segmentation from intraoral scanner images is crucial for computer-aided orthodontic surgical planning.
• Current deep learning methods struggle with effectively utilizing diverse geometric attributes like coordinates and normal vectors.
• Naive attribute concatenation in single-stream networks can lead to confusion and hinder learning of robust geometric representations.

Purpose of the Study:

• To develop a novel deep learning architecture, the two-stream graph convolutional network (TSGCN), for enhanced 3D tooth segmentation.
• To address the challenge of inter-view confusion arising from different raw geometric attributes.
• To improve the accuracy and automation of tooth segmentation in dental applications.

Main Methods:

• Designed a two-stream graph convolutional network (TSGCN) with input-specific graph-learning streams.
• Extracted complementary high-level geometric representations from coordinates and normal vectors separately.
• Employed a self-attention module to fuse single-view representations adaptively for discriminative multi-view learning.

Main Results:

• The TSGCN effectively handles inter-view confusion between different geometric attributes.
• The proposed method achieved significantly improved performance compared to state-of-the-art methods.
• Demonstrated superior accuracy in 3D tooth surface segmentation on a real-patient dataset.

Conclusions:

• The TSGCN offers a more effective approach to fusing complementary geometric information for tooth segmentation.
• This method enhances the learning of discriminative multi-view geometric representations.
• The TSGCN shows great potential for accurate and fully automatic tooth segmentation in orthodontic applications.