TSegLab: multi-stage 3D dental scan segmentation and labeling

Ahmed Rekik1, 2, Achraf Ben-Hamadou1, 2, Oussama Smaoui2, Firas Bouzguenda2, Sergi Pujades3, Edmond Boyer3,
1Centre de Recherche en Numerique de Sfax, Tunisia
2Udini, France
3Inria, Univ. Grenoble Alpes, CNRS, Grenoble INP, LJK, France

Abstract

In modern dentistry, computer-aided design (CAD) tools are becoming more popular due to their high level of accuracy in dental treatment planning. In particular, advanced intraoral scanners are now widely utilized in CAD systems because they produce accurate 3D surface models of the intraoral cavity. This advancement has liberated dentists from tedious and time-consuming tasks. With CAD, dentists can predict and optimize orthodontic and/or restorative treatment outcomes through convenient simulations involving teeth extraction, movement, deletion, and rearrangement. The precise and reliable automatic segmentation and labeling of teeth are crucial components in CAD systems. In this article, we propose a novel deep learning approach for 3D teeth scan segmentation and labeling. Our approach is divided into three main tasks: teeth localization, segmentation, and labeling. Firstly, we leverage recent advances in 2D object detection based on convolutional neural networks (CNNs) to achieve robust localization of visible teeth in the scan. The detected teeth candidates are then fed into a semantic segmentation network for fine teeth crown segmentation. To address the teeth labeling task, we design a novel graph neural network that models both the 3D shape appearance and spatial distribution of teeth in the jaw. We evaluate our method on the Teeth3DS dataset, which comprises 1800 intraoral 3D scans. Experimental results demonstrate that our method outperforms state-of-the-art techniques.

Press R to reset view.

Method Overview

The proposed segmentation pipeline consists of three main stages.

1- Coarse teeth detection: based on detecting teeth in the rendered three-channel 2D representation of the input scan using mask-RCNN.

2- Fine teeth segmentation in 2D harmonic parameter space to separate the crown from the gingiva and surrounding teeth.

3- Teeth labeling: To identify each segmented tooth in the input intraoral 3D scan based on GNNs.

Results

Comparisons with state-of-the-art methods on TLA, TSA, TIR, standard 3D segmentation evaluation metrics (DSC, OA), and running-time T in seconds.

Citing us

@article{REKIK2025109535,
title = {TSegLab: Multi-stage 3D dental scan segmentation and labeling},
journal = {Computers in Biology and Medicine},
volume = {185},
pages = {109535},
year = {2025},
issn = {0010-4825},
doi = {https://doi.org/10.1016/j.compbiomed.2024.109535},
url = {https://www.sciencedirect.com/science/article/pii/S0010482524016202},
author = {Ahmed Rekik and Achraf Ben-Hamadou and Oussama Smaoui and Firas Bouzguenda and Sergi Pujades and Edmond Boyer},
keywords = {Dental scan segmentation, Teeth classification, Teeth segmentation, Graph neural network, 3D intraoral scan, Teeth3DS},
}

@article{ben2022teeth3ds,
    title={{Teeth3Ds+: An Extended Benchmark for Intra-oral 3D Scans Analysis}},
    author={Ben-Hamadou, Achraf and Neifar, Nour and Rekik, Ahmed and Smaoui, Oussama and Bouzguenda, Firas and Pujades, Sergi and  Boyer, Edmond and Ladroit, Edouard},
    journal={arXiv preprint arXiv:2210.06094},
    year={2022}
}