Transformer based 3D tooth segmentation via point cloud region partition.

Area of Science:

• Computer-aided dentistry
• Medical imaging
• Machine learning

Background:

• Accurate 3D tooth segmentation is crucial for computer-aided dentistry.
• Existing Transformer methods struggle with global context due to memory and computational constraints.

Purpose of the Study:

• To propose PointRegion, a novel Transformer-based network for efficient 3D tooth segmentation.
• To enable processing of entire point clouds at a low cost while effectively modeling global context.

Main Methods:

• PointRegion interprets point clouds as tessellations of learnable regions.
• A RegionPartition module partitions the point cloud into regions and projects them as embeddings.
• An offset-attention based RegionEncoder models global context among regions.
• A novel point-to-region association mechanism replaces traditional convolutions for point-level segmentation.

Main Results:

• PointRegion processes large-scale point clouds with low computational cost and memory consumption.
• The method achieves effective 3D tooth segmentation by modeling global context.
• Extensive experiments demonstrate the effectiveness and superiority of the proposed method on a dental dataset.

Conclusions:

• PointRegion offers an efficient and effective solution for 3D tooth segmentation in computer-aided dentistry.
• The novel approach overcomes limitations of existing methods in modeling global context.
• The method shows significant promise for advancing dental applications.