Detecting Dental Landmarks from Intraoral 3D Scans: the 3DTeethLand challenge

TL;DR

The paper discusses the 3DTeethLand challenge for detecting dental landmarks from intraoral 3D scans, highlighting dataset creation, evaluation benchmarks, and top-performing deep learning methods.

Contribution

It introduces a new benchmark dataset and challenge for 3D dental landmark detection, fostering methodological advances in this clinical application.

Findings

A publicly available dataset with 340 intraoral scans was released.

Top teams achieved high precision and balanced recall using innovative deep learning strategies.

The winning approach used a two-stage Stratified Transformer with segmentation and clustering.

Abstract

Teeth landmark detection is a key task in modern orthodontics, supporting advanced diagnosis, personalized treatment planning, and effective monitoring of treatment progress. However, several significant challenges may arise due to the intricate geometry of individual teeth and the substantial variations observed across different individuals. To address these complexities, the development of advanced techniques, especially through the application of deep learning, is essential for the precise and reliable detection of 3D tooth landmarks. In this context, the 3DTeethLand challenge was held in conjunction with the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) in 2024, calling for algorithms focused on teeth landmark detection from intraoral 3D scans. This challenge introduced a publicly available dataset for 3D dental landmark detection from 340 intraoral scans, providing a standardized benchmark to evaluate state-of-the-art approaches and encouraging methodological advances toward addressing this clinically problem. A total of 49 teams participated, and 6 teams reached the final phase. The winning team achieved a rank score of 0.91, with a mean Average Precision of 0.78 and a mean Average Recall of 0.65, demonstrating a balance between precision and recall. Top teams achieved high precision with different strategies: the first-ranked team used a two-stage Stratified Transformer with segmentation and weighted DBSCAN, while the second-ranked team adopted a single-stage DGCNN with offset regression and class-specific non-maximum suppression.