# TFormer: 3D Tooth Segmentation in Mesh Scans with Geometry Guided Transformer

**Authors:** Huimin Xiong, Kunle Li, Kaiyuan Tan, Yang Feng, Joey Tianyi Zhou, Jin Hao, Zuozhu Liu  
**arXiv:** [2210.16627](https://arxiv.org/abs/2210.16627)  
**Date:** 2022-10-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.16627v1/

## Key points

- Optical Intra-oral Scanners (IOS) are widely used in digital dentistry to provide 3D information of dental crowns and gingiva.
- Accurate segmentation of teeth and gingiva in IOS scans is crucial for dental applications.
- Previous segmentation methods have limitations in accurately delineating complex tooth-tooth or tooth-gingiva boundaries.
- The proposed method called TFormer is based on 3D transformer architectures and addresses these challenges.
- TFormer leverages local and global dependencies among different teeth to distinguish various types of teeth with diverse anatomical structures and challenging boundaries.
- The method introduces a geometry guided loss based on point curvature to refine boundary predictions for more accurate segmentation.
- A multi-task learning scheme is employed by introducing an additional teeth-gingiva segmentation head to improve overall performance.
- TFormer surpasses existing state-of-the-art baselines by achieving impressive accuracy, mean intersection over union (mIoU), and dice similarity coefficient (DSC) scores of 97.97%, 94.34%, and 96.01% respectively.
- The main contributions include the design of a 3D Transformer-based architecture, the introduction of a geometry guided loss, and the adoption of a multi-task learning scheme.
- A large-scale, high-resolution, and heterogeneous 3D IOS dataset was collected for comprehensive evaluation and clinical applicability testing.
- Previous methods for 3D intra oral tooth segmentation can be categorized into conventional methods with deep learning and deep learning-based methods.
- TFormer achieves state-of-the-art performance in 3D tooth segmentation on this large-scale dataset and exhibits great potential for real-world clinical applications.

## Summary

In the field of digital dentistry, Optical Intra-oral Scanners (IOS) are widely used to provide high-resolution 3D information of dental crowns and the gingiva. Accurate segmentation of teeth and gingiva in IOS scans is crucial for various dental applications. However, previous segmentation methods have shown limitations in accurately delineating complex tooth-tooth or tooth-gingiva boundaries, leading to unsatisfactory results across different patients. Additionally, these methods have not been validated with large-scale datasets. To address these challenges, this paper proposes a novel method called TFormer based on 3D transformer architectures. The method is evaluated using a large-scale and high-resolution 3D IOS dataset consisting of 16,000 complex dental models, making it the largest IOS dataset to date. TFormer leverages both local and global dependencies among different teeth to distinguish various types of teeth with diverse anatomical structures and challenging boundaries. The proposed method also introduces a geometry guided loss based on a novel point curvature to exploit boundary geometric features. This helps refine boundary predictions for more accurate and smooth segmentation. Furthermore, a multi-task learning scheme is employed by introducing an additional teeth-gingiva segmentation head to improve overall performance. Extensive experiments conducted on the large-scale dataset demonstrate that TFormer surpasses existing state-of-the-art baselines by a significant margin. The model achieves impressive accuracy, mean intersection over union (mIoU), and dice similarity coefficient (DSC) scores of 97.97%, 94.34%, and 96.01% respectively. The main contributions of this work include the design of a 3D Transformer-based architecture that effectively captures contextual information for distinguishing different types of teeth with complex anatomical structures and confusing boundaries. The introduction of a geometry guided loss based on point curvature enhances boundary refinement for more clinically applicable tooth segmentation. Additionally, the adoption of a multi-task learning scheme improves overall segmentation performance. The paper also highlights the collection of a large-scale, high resolution, and heterogeneous 3D IOS dataset for comprehensive evaluation as well as its clinical applicability test which confirms its utility in real world scenarios.. In terms of related work, previous methods for 3D intra oral tooth segmentation can be categorized into conventional methods with deep learning and deep learning based methods; conventional methods often involve projecting the 3D mesh to 2D images before applying traditional image processing techniques for segmentation while deep learning based methods leverage neural networks to directly process raw 3D data for more accurate segmentations.. The experimental results demonstrate that TFormer achieves state -of -the -art performance in 3D tooth segmentation on this large scale dataset and exhibits great potential for real world clinical applications .The rest of the paper is organized as follows: Section 2 provides brief review related work including 3 D tooth segmentation methods ,3 D geometric data semantic segmentation methods ,and transformer based approaches point clouds .Section three describes proposed approach detail .Implementation details ,comparison competing methods ,ablation studies ,visualization results ,advantages ,and limitations are presented section four .Finally paper concludes section five .