SemiTooth: a Generalizable Semi-supervised Framework for Multi-Source Tooth Segmentation

Conference: CVPR 2026 arXiv: 2603.11616 Code: N/A Area: Medical Image Segmentation Keywords: Semi-supervised Learning, Multi-Source Data, Tooth Segmentation, CBCT, Pseudo Labels

TL;DR

This paper proposes SemiTooth, a multi-teacher multi-student semi-supervised framework coupled with a Stricter Weighted-Confidence (SWC) constraint, which effectively leverages multi-source unlabeled data for multi-source CBCT tooth segmentation and achieves cross-source generalization.

Background & Motivation

CBCT (Cone Beam Computed Tomography) tooth segmentation is a core task in clinical dental diagnosis, yet it faces two major challenges:

1. Scarcity of annotated data: Voxel-level annotation is extremely costly, leaving large volumes of de-identified CBCT data unutilized.
2. Multi-source data heterogeneity: CBCT data acquired from different institutions and devices exhibit significant distribution discrepancies (in density distribution, grayscale intensity, and feature clustering), making unified model training difficult.

Existing semi-supervised medical image segmentation methods (e.g., Mean Teacher, Co-training) are predominantly designed for single-source data and cannot effectively handle distribution shifts in multi-source settings. Furthermore, publicly available multi-source CBCT tooth segmentation datasets are extremely scarce, limiting method validation.

Method

Overall Architecture

SemiTooth is a multi-branch semi-supervised framework adopting a three-student two-teacher architecture:

• Multi-source data are reorganized into three subsets: main (labeled primary source), other (unlabeled data from other sources), and mixed (an unlabeled subset with distribution similar to the primary source, selected via Wasserstein distance measurement).
• Three student networks are each responsible for learning from one subset.
• Two teacher networks supervise the students on the mixed and other subsets respectively, providing stable pseudo labels through EMA updates.

Key Designs

1. Multi-source subset partitioning strategy:
2. Inter-source Wasserstein distances are used to divide unlabeled data into a mixed group (distribution close to the labeled source) and an other group (larger distribution discrepancy).

3. The mixed subset serves as a bridge across different sources, improving cross-source training robustness.

4. Multi-Teacher Multi-Student architecture:

5. Three student networks share similar architectures but learn independently, promoting effective knowledge transfer while maintaining diversity.
6. Teachers are updated via EMA: \(\theta_t^{(k)} \leftarrow \gamma \theta_t^{(k-1)} + (1-\gamma) \theta_s^{(k)}\), with decay rate \(\gamma = 0.99\).

7. Compared to Mean Teacher (single teacher, single student) and Co-training (no teacher, multiple students), SemiTooth combines the stability of teacher supervision with the cross-source capability of multi-student collaboration.

8. Stricter Weighted-Confidence (SWC) Constraint:

9. Each sample is uniformly partitioned into non-overlapping cubic regions \(\{r\}\).
10. Region-level confidence is defined as the mean of the maximum class probability across all voxels within a region: \(c(r) = \mathbb{E}_{i \in r}[\max_c P_{i,c}^T]\).
11. Low-confidence regions (\(c(r) < \tau\)) are flagged as unreliable and ignored.
12. Within reliable regions, voxel-level confidence \(c_i = \max_c P_{i,c}^T\) further weights teacher–student alignment.
13. This dual-layer design balances structural reliability at the region level with fine-grained accuracy at the voxel level.

Loss & Training

The total loss consists of three components:

• Supervised loss (labeled data from the primary source): \(\mathcal{L}_{sup} = CE(P^S(x^l), y)\)
• SWC consistency loss (unlabeled data from other and mixed sources):

\[\mathcal{L}_{SWC} = \mathbb{E}_{r \in \mathcal{R}_\tau} \left[ \mathbb{E}_{i \in r} \left[ c_i \cdot CE(P_i^S, P_i^T) \right] \right]\]

• Total loss: \(\mathcal{L}_{total} = \mathcal{L}_{sup} + \alpha \mathcal{L}_{cons}^u + \beta \mathcal{L}_{cons}^h\)

where \(\alpha = \beta = 0.5\) balance contributions from different sources, and the confidence threshold is set to \(\tau = 0.9\).

Training details: V-Net is used as the backbone, with the Adam optimizer, learning rate \(10^{-4}\), batch size 4, trained for 300 epochs on 4 NVIDIA A4500 GPUs.

Key Experimental Results

Dataset: MS3Toothset

A self-constructed multi-source semi-supervised tooth segmentation dataset comprising 98 annotated samples (20 for testing) and 438 unlabeled samples from three sources (ShanghaiTech, PKU-SS, and AFMC). The sources exhibit significant differences in density, grayscale intensity, and feature distributions.

Main Results

Method	Venue	Year	mIoU	Dice	Recall	Acc
V-Net	IEEE 3DV	2016	61.36	73.65	70.77	66.75
Mean Teacher	NeurIPS	2017	67.69	78.72	78.06	73.68
UA-MT	MICCAI	2019	68.37	79.18	80.42	76.17
ASDA	IEEE TIP	2022	73.75	83.63	80.93	78.79
MLRPL	MIA	2024	72.86	83.29	79.75	77.39
CMT	ACM MM	2024	76.14	85.07	87.14	84.32
Uni-HSSL	CVPR	2025	75.76	85.42	84.26	81.88
SemiTooth	-	2025	76.67	85.69	88.66	86.44

Ablation Study

Exp	V-Net	MT	ST	SWC	mIoU	Dice	Recall	Acc
1	✓				61.36	73.65	70.77	66.75
2	✓	✓			67.69	78.72	78.06	73.68
3	✓	✓		✓	69.94	80.29	79.67	75.34
4	✓	✓	✓		75.37	84.56	83.07	80.48
5	✓	✓	✓	✓	76.67	85.69	88.66	86.44

Key Findings

1. The SemiTooth multi-branch architecture (Exp4 vs. Exp2) contributes the largest performance gain (+7.68 mIoU), demonstrating that the multi-teacher multi-student structure is critical for cross-source learning.
2. The SWC constraint yields +2.25 mIoU over the Mean Teacher baseline (Exp3 vs. Exp2) and +1.30 mIoU on top of SemiTooth (Exp5 vs. Exp4).
3. t-SNE visualizations show that after training with SemiTooth, feature distributions across different sources become more clustered, validating the improvement in cross-source generalization.

Highlights & Insights

• Clear problem formulation: This work is the first to systematically address multi-source semi-supervised CBCT tooth segmentation and constructs a dedicated benchmark dataset.
• Well-motivated architecture: The asymmetric three-student two-teacher design is more effective than a naive multi-student framework, as the teachers provide stable pseudo-label training signals.
• Elegant SWC constraint: The dual-layer filtering strategy—region-level gating combined with voxel-level weighting—is better suited to the spatial structure of 3D CBCT than simple confidence thresholding.
• Wasserstein-based subset partitioning: Using distributional distance to bridge labeled and unlabeled sources is a practical and effective strategy.

Limitations & Future Work

1. Limited dataset scale: Only 98 annotated and 438 unlabeled samples; validation on larger-scale data is needed.
2. Restricted number of sources: Only three sources are considered; scalability to a larger number of sources has not been verified.
3. Single backbone: Only V-Net is evaluated; more advanced 3D segmentation networks (e.g., nnU-Net, SwinUNETR) may yield further improvements.
4. Out-of-distribution generalization unverified: Model performance on completely unseen new-source data remains unknown.
5. Fixed cubic region size in SWC: An adaptive region partitioning scheme may be preferable.

Related Work & Insights

• The Mean Teacher paradigm is the foundational framework for semi-supervised medical image segmentation (SSMIS); SemiTooth extends it via a multi-branch design.
• CMT (ACM MM 2024) and Uni-HSSL (CVPR 2025) are recent multi-source semi-supervised methods, both of which SemiTooth outperforms.
• Cross-domain medical segmentation provides inspiration: Wasserstein distance for source distribution measurement and multi-teacher structures for cross-domain knowledge transfer.

Rating

Dimension	Score
Novelty	⭐⭐⭐
Theoretical Depth	⭐⭐⭐
Experimental Thoroughness	⭐⭐⭐
Value	⭐⭐⭐⭐
Writing Quality	⭐⭐⭐
Overall	⭐⭐⭐

Related Papers

• [CVPR 2026] Addressing Data Scarcity in 3D Trauma Detection through Self-Supervised and Semi-Supervised Learning with Vertex Relative Position Encoding
• [CVPR 2026] A Semi-Supervised Framework for Breast Ultrasound Segmentation with Training-Free Pseudo-Label Generation and Label Refinement
• [CVPR 2026] Uncertainty-Aware Concept and Motion Segmentation for Semi-Supervised Angiography Videos
• [CVPR 2026] Semantic Class Distribution Learning for Debiasing Semi-Supervised Medical Image Segmentation
• [CVPR 2026] SCDL: Semantic Class Distribution Learning for Debiasing Semi-Supervised Medical Image Segmentation