Geometric-Guided Few-Shot Dental Landmark Detection with Human-Centric Foundation Model

Anbang Wang, Marawan Elbatel, Keyuan Liu, Lizhuo Lin, Meng Lan, Yanqi Yang, Xiaomeng Li

Introduction

Precision in orthodontics and periodontics is critical for bone health assessment and implant planning. Identifying landmarks such as the root apex or cementoenamel junction (CEJ) on Cone-Beam Computed Tomography (CBCT) scans is currently a manual, labor-intensive process prone to significant inter-expert variability. While deep learning offers automation, the scarcity of expert-annotated dental radiology data poses a major challenge. Existing few-shot methods often fail to achieve the 0.5 mm clinical precision standard required for surgical applications.

GeoSapiens addresses this gap by leveraging a human-centric foundation model and anatomical geometric constraints to achieve high-precision detection with minimal training data.

Methodology

The GeoSapiens framework integrates three core innovations to overcome data scarcity:

  1. Human-Centric Foundation Model: The system utilizes Sapiens, a vision transformer (ViT) pre-trained on 300 million human images. The researchers hypothesized that the spatial relationships and symmetries learned from human body poses provide a robust prior for the symmetrical structures of human teeth.
  2. Parameter-Efficient Fine-Tuning (LoRA): To prevent overfitting on sparse datasets, the team employed Low-Rank Adaptation (LoRA). This allowed them to train only 0.3% of the model's parameters (reducing trainable parameters from 330 million to 24 million), ensuring a more stable and efficient training process.
  3. Geometric Loss ($L_{geo}$): This novel loss function enforces anatomical rules by penalizing predictions that violate known dental structures. Specifically, it ensures the predicted tooth axis is perpendicular to root levels and that root-level lines remain parallel. A Soft-argmax operation converts heatmaps into differentiable coordinates to enable this geometric supervision.
The GeoSapiens architecture, featuring the Sapiens backbone with LoRA and the geometric loss integration.

Experimental Setup

The framework was evaluated using a new benchmark dataset, LDTeeth, which contains 347 lateral CBCT slices with 16 landmarks per tooth. In a rigorous few-shot setting, the model was trained on images from only three patients (36 images total) and tested on 162 images from 16 different patients. Performance was measured using the Success Detection Rate (SDR) at various thresholds and Mean Radial Error (MRE).

Results and Discussion

GeoSapiens significantly outperformed state-of-the-art (SOTA) landmark detection models, including GU2Net, FM-OSD, and NFDP.

Quantitative comparison of GeoSapiens against baseline models across various SDR thresholds and MRE.

Key results include:

  • Clinical Precision: At the strict 0.5 mm SDR threshold, GeoSapiens achieved 63.19%, an 8.18% improvement over the next best model (NFDP).
  • Error Reduction: The framework reduced the Mean Radial Error to 0.747 mm, nearly halving the error seen in several other few-shot approaches (1.3–1.5 mm).
  • Ablation Insights: The addition of geometric loss alone improved 0.5 mm accuracy by 2.5% compared to the baseline, proving that explicit anatomical "cheat sheets" are highly effective in low-data regimes.

Conclusion

GeoSapiens demonstrates that cross-domain synergy between human-centric foundation models and specialized dental anatomy can overcome the hurdles of medical data scarcity. While currently focused on 2D slices of anterior teeth, future work will extend these geometric constraints to complex multi-root molar structures and full 3D CBCT volumes to further enhance clinical utility.

Summary updated

BibTeX
@inproceedings{wang2025geometric,
  author = {Anbang Wang and Marawan Elbatel and Keyuan Liu and Lizhuo Lin and Meng Lan and Yanqi Yang and Xiaomeng Li},
  title = {{Geometric-Guided Few-Shot Dental Landmark Detection with Human-Centric Foundation Model}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2025},
  doi = {10.1007/978-3-032-04971-1\_19},
}
Figure