MICCAI STSR 2025 Challenge: Semi-Supervised Teeth and Pulp Segmentation and CBCT-IOS Registration

TL;DR

The paper presents STSR 2025, a MICCAI satellite challenge benchmarking semi-supervised methods for 3D CBCT teeth and pulp canal segmentation and CBCT–IOS crown-root registration. It introduces two clinically motivated tasks and a public, multi-center dataset with labeled/unlabeled data, plus rigorous, multi-metric evaluation and Docker-based reproducibility. Key findings show SSL substantially boosts 3D segmentation accuracy and enables robust cross-modal registration, with state-space and transformer-inspired hybrids (e.g., Mamba CMamba blocks, TAAM) offering the strongest gains in certain settings. The work demonstrates meaningful clinical impact through human–machine annotation studies, revealing large time savings and potential for scalable dental AI deployment, while outlining future directions to enhance generalizability and efficiency.

Abstract

Cone-Beam Computed Tomography (CBCT) and Intraoral Scanning (IOS) are essential for digital dentistry, but annotated data scarcity limits automated solutions for pulp canal segmentation and cross-modal registration. To benchmark semi-supervised learning (SSL) in this domain, we organized the STSR 2025 Challenge at MICCAI 2025, featuring two tasks: (1) semi-supervised segmentation of teeth and pulp canals in CBCT, and (2) semi-supervised rigid registration of CBCT and IOS. We provided 60 labeled and 640 unlabeled IOS samples, plus 30 labeled and 250 unlabeled CBCT scans with varying resolutions and fields of view. The challenge attracted strong community participation, with top teams submitting open-source deep learning-based SSL solutions. For segmentation, leading methods used nnU-Net and Mamba-like State Space Models with pseudo-labeling and consistency regularization, achieving a Dice score of 0.967 and Instance Affinity of 0.738 on the hidden test set. For registration, effective approaches combined PointNetLK with differentiable SVD and geometric augmentation to handle modality gaps; hybrid neural-classical refinement enabled accurate alignment despite limited labels. All data and code are publicly available at https://github.com/ricoleehduu/STS-Challenge-2025 to ensure reproducibility.

Figures (13)

• Figure 1: Overview of the Semi-Supervised Teeth Segmentation and Registration (STSR 2025) framework and its clinical utility. The workflow proceeds from (Left) multi-modal data acquisition, capturing volumetric Cone-Beam Computed Tomography (CBCT) and surface-based Intraoral Scans (IOS). These inputs are processed in (Middle) the core semi-supervised learning paradigm, utilizing Encoder-Decoder architectures (e.g., U-Mamba, PointNet++) to perform two key tasks: high-precision segmentation of teeth and root pulp canals, and cross-modal rigid registration of IOS crowns with CBCT roots. The resulting 3D models facilitate (Right) diverse downstream clinical applications, including orthodontic treatment planning, endodontic (root canal) therapy, implant surgery simulation, and Large Language Model (LLM)-driven automated diagnostic reporting.
• Figure 2: End-to-end workflow of the MICCAI 2024 Semi-supervised Teeth Segmentation (STS) Challenge. The process encompasses five key stages: (1) multi-center data collection to ensure dataset diversity, (2) iterative annotation by clinicians for high-quality ground truth, (3) construction of the semi-supervised dataset with distinct labeled and unlabeled sets, and (4/5) the final summarization and evaluation of submitted participant methods.
• Figure 3: Overview of the STS 2025 Challenge participation and timeline. (a) World map illustrating the geographical distribution of registered teams. (b) Detailed timeline of the challenge schedule from registration and training to the final workshop and announcement of results, together with the number of participating teams (78 for Task 1 and 52 for Task 2) and the total submissions (392 for Task 1 and 29 for Task 2).
• Figure 4: Overview of prominent methodological strategies employed by participants in the STS 2024 Challenge. The figure illustrates four key approaches: (a) Knowledge transfer with pretrained models, where pre-trained foundation models (e.g., SAM) are leveraged to improve segmentation. (b) Consistency regularization learning, including self-teaching, model perturbation, and semantic consistency ($\mathcal{T}_{1}$ and $\mathcal{T}_{2}$ denote two kinds of transformation). (c) Multi-stage architecture optimization decomposes the problem into multiple sub-problems and gradually obtained fine results.
• Figure 5: Overall diagram of the proposed U-Mamba2-SSL framework. (a) UMamba2 is first pre-trained by reconstructing inputs corrupted with noise, downsampled, and masked; (b) The second stage involves a combination of supervisedloss for the labeled data and consistency regularization between the unperturbedoutput and the perturbed output of the unlabeled data; (c) The final stage introduces pseudo labeling on top of the training objectives in (b). Only pseudo labels with confidence above a certain threshold contribute to the training loss.