Abstract
Accurate 3D tooth segmentation is essential for digital dentistry but suffers from the high cost of expert annotation. Semi-supervised learning (SSL) offers a solution by leveraging unlabeled data, yet it is often hindered by noisy pseudo-labels. We propose GeoT, a novel framework that explicitly models label noise using an Instance-dependent Transition Matrix (IDTM). By anchoring this noise model with point-level and class-level geometric priors, GeoT effectively rectifies supervision signals. Our approach achieves performance comparable to fully supervised models while using only 20% of the labeled data on the Teeth3DS benchmark.
Introduction
High-quality 3D point cloud annotations for intra-oral scans (IOS) require significant clinical expertise and manual labor. While semi-supervised learning leverages small labeled sets alongside large unlabeled datasets, the resulting "pseudo-labels" are frequently incorrect. Traditional SSL methods often discard low-confidence labels, wasting potentially valuable information. While Noise Transition Matrices (NTMs) can theoretically model this noise, applying them to dense 3D point clouds is computationally prohibitive because every point requires its own noise model. GeoT addresses this by using geometric constraints to tame the complexity of point-wise noise modeling.
Methodology
GeoT introduces a structured approach to noise modeling through three core components:
- Instance-dependent Transition Matrix (IDTM): Instead of a global noise model, GeoT estimates a unique transition matrix for every point, representing the probability of a "clean" class being mislabeled as a "noisy" one.
- Point-level Geometric Regularization (PLGR): This module enforces spatial smoothness by assuming that adjacent points of the same tooth category should exhibit similar noise patterns. By minimizing the variance between neighboring NTMs, it simplifies the optimization landscape.
- Class-level Geometric Smoothing (CLGS): This leverages the FDI World Dental Federation notation—the standard anatomical map of human teeth. Since teeth follow a fixed spatial order, a Gaussian prior is used to suggest that a point is more likely to be confused with its immediate anatomical neighbors than with distant teeth.
Experimental Results
We evaluated GeoT on the Teeth3DS public benchmark (1,800 scans) and a private dataset of ,1200 unlabeled scans. GeoT consistently outperformed state-of-the-art SSL methods such as FixMatch and UA-MT across various labeling ratios.
Key Performance Metrics
- Data Efficiency: With only 20% labeled data, GeoT achieved a mean Intersection over Union (mIoU) of 83.83%, nearly matching the fully supervised upper bound of 84.31%.
- Low-Data Robustness: In an ultra-low data regime (0.5% labels, or roughly 6 scans), GeoT maintained a 56.81% mIoU, significantly higher than competing methods.
- Qualitative Quality: GeoT produces cleaner boundaries between adjacent teeth, effectively preventing the "bleeding" of classes often seen in other models.
Conclusion and Future Work
GeoT demonstrates that modeling label noise through geometric priors is a powerful alternative to simple confidence filtering in SSL. By treating anatomy as a mathematical constraint, the framework makes dense point-wise noise estimation feasible. While the current model relies on standard dental arrangements, future work will focus on adapting the CLGS module to handle clinical outliers, such as missing teeth or severe overcrowding, where standard spatial priors may be violated.
Summary updated