Geodesic heatmap-based segmentation-free framework for robust tooth landmark detection

Automatic tooth landmark detection is a crucial component in orthodontic treatment, aiding in tooth morphology assessment, treatment planning, and oral health monitoring. However, challenges remain due to diverse landmark types, varying landmark quantities, anatomical variations, and dental abnormalities. To address these issues, this paper proposes a robust two-stage framework for precise landmark detection. In the first stage, an adaptive partitioning strategy employs a lightweight network to predict tooth centroids, which are then used to partition the dental points into localized patches, eliminating the need for precise segmentation. In the second stage, a geodesic distance-based heatmap is introduced to improve landmark detection accuracy. Furthermore, an anatomy-aware spatial augmentation strategy is proposed to simulate clinically challenging scenarios, thereby improving the model’s learning capability and its robustness, particularly in cases of abnormal teeth. Extensive experimental results on a public dataset demonstrate the superiority of our method, with significant improvements over state-of-the-art approaches.