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

Conference: CVPR 2026 arXiv: 2603.11616 Code: N/A Area: Medical Imaging Keywords: tooth segmentation, CBCT, semi-supervised learning, multi-source data, multi-teacher multi-student

TL;DR

This paper proposes SemiTooth, a framework that addresses annotation scarcity and cross-source domain discrepancy in multi-source CBCT tooth segmentation via a multi-teacher multi-student architecture and Strict Weighted Confidence (SWC) constraints. It also introduces MS3Toothset, the first multi-source semi-supervised tooth segmentation dataset.

Background & Motivation

CBCT tooth structure segmentation is central to intelligent oral diagnosis, yet faces two major challenges:

Annotation scarcity: Voxel-level annotation is costly and time-consuming, leaving large quantities of de-identified CBCT data unutilized.

Multi-source domain discrepancy: CBCT data from different institutions and devices exhibit significant differences in density distribution, intensity distribution, and feature space (verified via Kernel Density Estimation and t-SNE visualization), making cross-source generalization difficult.

Existing semi-supervised medical segmentation methods (Mean Teacher, UA-MT, MCF, etc.) are primarily designed for single-source data and lack cross-source knowledge transfer capability. Multi-source methods such as ASDA and Dual-Teacher either require complex networks or strong supervision, or have not been validated on multi-source CBCT tooth data.

Method

Overall Architecture

SemiTooth adopts a three-student + two-teacher multi-branch architecture. Data are organized into three subsets: - Main (labeled data from the primary source) - Other (unlabeled data from other sources) - Mixed (unlabeled data with distribution similar to the primary source, selected via Wasserstein distance)

Each subset is processed by a corresponding student network. Two teachers supervise the Mixed and Other students respectively, updated via EMA:

\[\theta_t^{(k)} \leftarrow \gamma \theta_t^{(k-1)} + (1-\gamma) \theta_s^{(k)}, \quad \gamma = 0.99\]

Key Designs

1. Multi-Teacher Multi-Student Architecture (SemiTooth):

   • Compared to Mean Teacher (single teacher–single student): SemiTooth assigns dedicated students to data from different sources, with teachers providing cross-source knowledge guidance.
   • Compared to Co-training (multi-student with shared weights, no teacher): SemiTooth provides stable pseudo-labels via EMA teachers.
   • The Mixed subset serves as an inter-source bridge, containing unlabeled samples with distributions similar to the primary source, alleviating training instability from direct cross-source learning.
   • Student networks share similar architectures to facilitate knowledge transfer while maintaining sufficient diversity.

2. Strict Weighted Confidence Constraint (SWC):

Addresses the degradation of consistency regularization caused by noise introduced by CBCT heterogeneity. SWC combines region-level gating with voxel-level weighting:

- **Region-level gating**: Samples are uniformly divided into non-overlapping cubic regions $\{r\}$. Region confidence is computed as $c(r) = \mathbb{E}_{i \in r}[\max_c P^T_{i,c}]$; regions below threshold $\tau=0.9$ are deemed unreliable and discarded.
- **Voxel-level weighting**: Within retained regions, voxel-level confidence $c_i = \max_c P^T_{i,c}$ is used to weight the teacher–student alignment:
$\mathcal{SWC}(P^S, P^T) = \mathbb{E}_{r \in \mathcal{R}_\tau}\left[\mathbb{E}_{i \in r}\left[c_i \cdot \mathcal{A}(P^S_i, P^T_i)\right]\right]$
- This design is particularly suited to 3D CBCT data, as region-level filtering exploits the spatial continuity inherent in volumetric data.

1. Multi-Source Dataset MS3Toothset:

   • Collected from three sources: ShanghaiTech (public, labeled), PKU-SS, and AFMC (private, unlabeled).
   • After preprocessing and filtering: 98 labeled samples (20 for testing) and 438 unlabeled samples.
   • The first comprehensive dataset for multi-source semi-supervised tooth segmentation.

Loss & Training

The total loss combines a supervised loss and two SWC consistency losses:

\[\mathcal{L}_{total} = \mathcal{L}_{sup} + \alpha \mathcal{L}_{cons}^u + \beta \mathcal{L}_{cons}^h, \quad \alpha = \beta = 0.5\]

• \(\mathcal{L}_{sup} = \text{CE}(P^S(x^l), y)\): supervised loss on labeled data from the primary source.
• \(\mathcal{L}_{cons}^u\), \(\mathcal{L}_{cons}^h\): SWC losses for the Other and Mixed sources, respectively.
• Backbone: V-Net; optimizer: Adam; learning rate: 0.0001; training: 300 epochs.

Key Experimental Results

Main Results

Method	mIoU	Dice	Recall	Acc
V-Net (fully supervised baseline)	61.36	73.65	70.77	66.75
Mean Teacher	67.69	78.72	78.06	73.68
UA-MT	68.37	79.18	80.42	76.17
ASDA	73.75	83.63	80.93	78.79
CMT	76.14	85.07	87.14	84.32
Uni-HSSL	75.76	85.42	84.26	81.88
SemiTooth	76.67	85.69	88.66	86.44

Ablation Study

Exp	Configuration	mIoU	Dice	Recall	Acc
1	V-Net	61.36	73.65	70.77	66.75
2	+ Mean Teacher	67.69	78.72	78.06	73.68
3	+ SWC (w/o SemiTooth)	69.94	80.29	79.67	75.34
4	+ SemiTooth (w/o SWC)	75.37	84.56	83.07	80.48
5	+ SemiTooth + SWC	76.67	85.69	88.66	86.44

Key Findings

• The multi-branch SemiTooth architecture contributes the most (Exp 2→4: mIoU +7.68), demonstrating the importance of source-aware learning for multi-source data.
• SWC yields consistent improvements under both single-teacher (Exp 2→3: +2.25 mIoU) and multi-teacher (Exp 4→5: +1.30 mIoU) settings.
• SemiTooth's advantage is most pronounced on the Recall metric (88.66 vs. second-best 87.14), which is clinically meaningful for reducing missed detection rates.
• t-SNE visualizations confirm that SemiTooth effectively reduces feature distribution gaps across sources.

Highlights & Insights

• Dataset contribution: MS3Toothset fills a gap in multi-source semi-supervised tooth segmentation benchmarks.
• The two-stage region-level + voxel-level filtering in SWC is intuitively grounded and well-suited to the spatial continuity of 3D medical data.
• Using Wasserstein distance to select distributionally similar samples for the Mixed subset as an inter-source bridge is a simple yet effective design choice.
• The ablation study is well-structured, clearly illustrating both the individual and joint contributions of each component.

Limitations & Future Work

• MS3Toothset is relatively small in scale (98 labeled + 438 unlabeled samples).
• Validation is limited to the authors' own dataset; generalizability to public standard benchmarks remains unknown.
• The teacher count and subset partitioning strategy depend on the choice of Wasserstein distance threshold, with no sensitivity analysis provided.
• Only V-Net is used as the backbone; effectiveness with stronger backbones (e.g., nnUNet, Swin UNETR) is not evaluated.
• Ablation of the region size and threshold \(\tau\) in SWC is insufficient.

Related Work & Insights

• SemiTooth outperforms recent multi-source methods such as CMT (ACM MM 2024) and Uni-HSSL (CVPR 2025).
• The multi-teacher multi-student paradigm is generalizable to other multi-domain semi-supervised scenarios (e.g., multi-center CT/MRI analysis).
• The SWC constraint can be readily integrated into any Mean Teacher-based framework.

Rating

• Novelty: ⭐⭐⭐ The multi-teacher multi-student and confidence-weighting ideas are not entirely new, but their systematic integration for tooth segmentation is valuable.
• Experimental Thoroughness: ⭐⭐⭐ Ablations are complete, but validation is limited to a single in-house dataset without external benchmarking.
• Writing Quality: ⭐⭐⭐ The framework is described clearly, though the paper is short (ICASSP length) and lacks detail in places.
• Value: ⭐⭐⭐ The dataset contribution has clinical value; generalizability of the method remains to be verified.

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