ToothFairy3 Challenge: 3D Dental CBCT Segmentation
- ToothFairy3 Challenge is a 3D dental CBCT segmentation benchmark targeting 46 dental and maxillofacial substructures through multi-anatomy and interactive tasks.
- It addresses complex issues such as anatomical scale variation, severe class imbalance, and CBCT artifacts using advanced multi-scale and anatomy-aware modeling.
- Evaluation metrics like Dice, HD95, and inference time drive the development of robust methods for precise segmentation in challenging dental imaging.
Searching arXiv for ToothFairy3 Challenge and closely related dental CBCT segmentation papers. The ToothFairy3 Challenge is a MICCAI 2025 benchmark for dental anatomy segmentation in cone-beam computed tomography (CBCT), centered on volumetric delineation of teeth and associated maxillofacial structures under substantial anatomical scale variation, class imbalance, and artifact burden. In the challenge setting described by direct participant reports, the dataset contains 532 CBCT scans annotated with 77 substructure classes, later consolidated into 46 substructures, and the benchmark includes a multi-anatomy segmentation task as well as an interactive segmentation task focused on the inferior alveolar nerves (LaBella et al., 18 Aug 2025, Tan et al., 15 Sep 2025). The challenge is notable for combining large structures such as mandible and maxilla with much smaller targets such as incisive nerves, lingual nerve, canals, pulp, and lingual foramen, thereby making resolution management, robustness, and anatomy-aware modeling central technical issues.
1. Benchmark definition and task scope
ToothFairy3 is fundamentally a 3D CBCT segmentation challenge rather than a 2D panoramic radiography benchmark or a tooth enumeration task. In the direct challenge descriptions, Task 1 is a multi-anatomy segmentation problem extending ToothFairy2 by adding pulp, incisive nerves, and lingual foramen, bringing the total to 46 anatomy classes, while Task 2 is an interactive segmentation task focused on the inferior alveolar nerves and uses user clicks as prompts (Tan et al., 15 Sep 2025). This immediately distinguishes ToothFairy3 from related dental AI tasks that emphasize tooth numbering, disease tagging, or panoramic-image boundary delineation.
The challenge is clinically motivated by the role of CBCT in diagnosis, treatment planning, and surgery. One reported ToothFairy3 solution emphasizes its relevance to radiation oncology, where accurate tooth-level and nerve-level segmentation can support dose mapping, dental risk assessment, and prevention of complications such as osteoradionecrosis, which disproportionately affects the mandible and is strongly dose-dependent (LaBella et al., 18 Aug 2025). Another challenge-specific report frames the same segmentation problem in terms of anatomically complex, noisy, artifact-prone, and symmetry-confounded CBCT volumes, with very small classes that are easily missed (Tan et al., 15 Sep 2025).
A common misconception is to treat ToothFairy3 as synonymous with generic tooth segmentation. The challenge, as described in the cited reports, is broader: it targets multi-structure dental and maxillofacial segmentation and includes an interactive nerve-segmentation setting, not merely voxel-wise separation of crowns or whole teeth (Tan et al., 15 Sep 2025).
2. Dataset composition and anatomical difficulty
The dataset composition reported for ToothFairy3 is unusually demanding. One direct challenge paper states that the dataset contains 532 CBCT scans annotated with 77 substructure classes, later consolidated by the organizers into 46 substructures (LaBella et al., 18 Aug 2025). The challenge data mix large, high-volume structures such as mandible, maxilla, and teeth with small structures such as incisive nerves, lingual nerve, canals, and pulp, and this imbalance is described as central to method design (LaBella et al., 18 Aug 2025).
Task 1 scan sizes are reported to range from to , which places substantial pressure on GPU memory and makes full-resolution training difficult without aggressive resampling or patch-based inference (Tan et al., 15 Sep 2025). A representative challenge submission used only 63 full-field-of-view “Set B” cases from the training data, noting that these scans had native spacing of about 0.3 mm isotropic and were broader than the cropped Set A and Set C scans, which often miss evaluated structures (LaBella et al., 18 Aug 2025). This indicates that not all available scans are equally suitable for full multi-structure supervision.
The anatomy mix also produces distinct failure regimes. Large structures can often be segmented with very high Dice, whereas tiny tubular or low-contrast structures remain difficult. In one ToothFairy3 solution, Phase 1 mandible Dice was about 0.977, maxilla Dice about 0.941, and pharynx Dice about 0.975, while incisive nerves and lingual nerve had very poor Phase 1 Dice, often near zero (LaBella et al., 18 Aug 2025). This structure-dependent disparity is one of the clearest empirical signatures of the challenge.
3. Evaluation priorities and challenge-specific objectives
The reported ToothFairy3 evaluation emphasizes overlap, boundary quality, and efficiency. For Task 1, one challenge paper states that evaluation includes Dice, HD95, and inference time, with inference speed explicitly part of the task evaluation (Tan et al., 15 Sep 2025). Another participant paper reports its performance on the ToothFairy3 out-of-sample validation set using average Dice and further reports structure-specific Dice behavior, especially for the difficult nerve classes (LaBella et al., 18 Aug 2025).
Task 2 introduces a different objective class: interactive segmentation. In the reported setup, the model receives a variable number of clicks, each with coordinates and a class label (Tan et al., 15 Sep 2025). This changes the problem from pure feed-forward segmentation to prompted refinement, and it makes feature fusion between volumetric image tokens and sparse user guidance a core design issue.
These evaluation priorities imply that ToothFairy3 is not satisfied by high region overlap alone. A plausible implication is that challenge-competitive methods must jointly manage large-anatomy coverage, small-structure recovery, and deployment-relevant runtime. This interpretation is reinforced by the fact that direct challenge submissions study tile size, mirroring axes, and post-processing not merely as engineering details but as factors that influence official evaluation (Tan et al., 15 Sep 2025).
4. Representative ToothFairy3 methods
The currently documented methodological landscape includes both direct challenge submissions and closely related dental segmentation approaches. The following papers are especially representative of the ToothFairy3 problem family.
| Paper | Relation to ToothFairy3 | Key reported result |
|---|---|---|
| DLaBella29 / Auto3DSeg + SegResNet (LaBella et al., 18 Aug 2025) | Direct MICCAI 2025 ToothFairy3 solution | Average Dice 0.87 on the out-of-sample validation set |
| U-Mamba2 (Tan et al., 15 Sep 2025) | Direct ToothFairy3 method for both tasks | Top 3 in both tasks; Task 1 mean Dice 0.792, Task 2 mean Dice 0.852 |
| YOLOrtho (Mei et al., 2023) | Related but not direct | End-to-end tooth enumeration and disease detection on panoramic X-rays, not ToothFairy3 |
One direct solution, the DLaBella29 system, uses a two-phase pipeline implemented in MONAI Auto3DSeg with a 3D SegResNet backbone (LaBella et al., 18 Aug 2025). Phase 1 segments full dental anatomy from full CBCT volumes after resampling to 0.6 mm isotropic resolution and intensity clipping to . The five fold-specific models are fused using Multi-Label STAPLE, and a second phase then crops tightly around the mandible, retains native 0.3 mm resolution, and focuses on the left incisive nerve, right incisive nerve, and lingual nerve (LaBella et al., 18 Aug 2025). The most important quantitative finding is that this second phase substantially improves the hardest nerve classes: lingual nerve from 0.0 to 0.681, left incisive nerve from 0.318 to 0.688, and right incisive nerve from 0.106 to 0.665 (LaBella et al., 18 Aug 2025).
Another direct method, U-Mamba2, is a hybrid CNN plus Mamba2 architecture built on nnU-Net and tailored explicitly to ToothFairy3 (Tan et al., 15 Sep 2025). Its design places Mamba2 only at the bottleneck of a U-Net-like encoder-decoder, uses self-supervised pretraining with the Disruptive Autoencoder framework on ToothFairy3 plus 371 unlabeled STS-3D-Tooth CBCT scans, and incorporates dental domain knowledge through label smoothing for related anatomies, class weighting of 10 for tiny classes, left-right mirroring with label swapping, and connected-component-based post-processing (Tan et al., 15 Sep 2025). It also includes an interactive branch for Task 2 in which click embeddings are fused with image features using two-way cross-attention blocks. On the hidden test data, the paper reports mean Dice 0.792 and HD95 93.19 for Task 1, and mean Dice 0.852 and HD95 7.39 for Task 2, securing a top 3 placement in both tasks (Tan et al., 15 Sep 2025).
YOLOrtho is frequently relevant in discussions of ToothFairy3 because it is a unified dental detector, but it is not a ToothFairy3 method. It is built for the Dentex Challenge 2023 and Tufts Dental data, frames diseases as attributes attached to their corresponding teeth, and targets tooth localization, enumeration, and disease detection in panoramic X-rays rather than CBCT segmentation (Mei et al., 2023). Its relevance is therefore conceptual rather than benchmark-specific.
5. Recurrent methodological themes
A clear technical theme in ToothFairy3 is multi-scale decomposition. The DLaBella29 submission explicitly separates full-volume anatomy segmentation from a high-resolution nerve-focused second phase, using mandible prediction as the spatial anchor for cropping (LaBella et al., 18 Aug 2025). U-Mamba2 instead concentrates global modeling at the bottleneck and supplements it with domain priors, self-supervised initialization, and symmetry-aware augmentation (Tan et al., 15 Sep 2025). Both approaches respond to the same constraint: a single undifferentiated pass is poorly matched to volumes containing both large jaw structures and very small tubular targets.
A second theme is explicit handling of symmetry and anatomy. U-Mamba2 notes that plain left-right mirroring can confuse the model, so mirrored training examples require label swapping, and mirrored test-time predictions also require swapping of left-right class logits (Tan et al., 15 Sep 2025). This is not merely augmentation hygiene; it reflects the fact that mirrored anatomies can be morphologically similar yet semantically distinct. The same paper also assigns class weight 10 to left incisive nerve, right incisive nerve, and lingual foramen, indicating that tiny-structure recovery must be enforced directly in the loss design (Tan et al., 15 Sep 2025).
A third theme is robustness through ensembling or consensus. In the DLaBella29 system, five-fold predictions are fused using Multi-Label STAPLE, which is described as producing cleaner borders than single-fold outputs, though it can suppress very small structures detected by only one model (LaBella et al., 18 Aug 2025). This trade-off is characteristic of ToothFairy3: consensus can regularize noisy outputs but may remove rare true positives, especially in small classes.
These patterns suggest that ToothFairy3 is best viewed not as a single-model contest but as a benchmark for architecture-plus-procedure design. The strongest reported systems combine backbone selection with scale-aware preprocessing, target-specific refinement, anatomy-aware supervision, and carefully tuned inference.
6. Position within the broader dental challenge ecosystem
ToothFairy3 sits within a broader sequence of dental imaging challenges that progressively expand anatomical scope and supervision difficulty. An earlier ToothFairy2023 paper focused on inferior alveolar nerve canal segmentation under mixed dense and sparse labels and converted the mixed-supervision problem into a semi-supervised one through connectivity-based selective re-training, achieving DSC 0.7956 and HD95 4.4905 on the final test set and winning the competition (Liu et al., 2023). This predecessor benchmark already established the importance of topology-aware pseudo-label selection for thin connected structures.
Semi-supervised teeth segmentation benchmarks provide another line of development. The STS MICCAI 2023 Challenge introduced semi-supervised tooth segmentation in both 2D panoramic X-rays and 3D CBCT, with only 12 labeled plus 300 unlabeled CBCT volumes in the final round, and highlighted multi-stage nnU-Net-centered pipelines, pseudo-label filtering, and frequency-domain augmentation as strong 3D strategies (Wang et al., 2024). The MICCAI STS 2024 Challenge then formalized semi-supervised instance-level tooth segmentation in panoramic X-ray and CBCT images, reporting that the 3D CBCT winner improved instance DSC from 30.80% for a fully supervised nnU-Net baseline to 92.15%, a gain of 61.35 points (Wang et al., 28 Nov 2025).
The STSR 2025 Challenge extends this trajectory beyond whole-tooth segmentation to teeth and pulp canal segmentation and CBCT-IOS registration, explicitly emphasizing that root pulp canals are curvilinear, extremely narrow, and exhibit low contrast against surrounding dentin (Wang et al., 2 Dec 2025). Taken together, these benchmarks suggest a broader shift from single-structure segmentation toward semi-supervised, multi-structure, and workflow-level dental AI. ToothFairy3 occupies a central position in that progression because it couples a large 3D CBCT benchmark with small-structure difficulty, anatomy count expansion, and an interactive task.
The challenge also clarifies what is not ToothFairy3. Panoramic-image methods such as YOLOrtho for end-to-end tooth enumeration and disease detection, or BFFNet for boundary-focused 2D tooth segmentation, address adjacent problems but operate in different modalities and with different endpoints (Mei et al., 2023, Zhang et al., 2024). Their value lies in transferable ideas—attribute modeling, boundary emphasis, or anatomy-aware post-processing—not in direct benchmark comparability.
In this sense, ToothFairy3 functions as both a competition and a technical stress test for current 3D dental segmentation methodology. Its reported solutions show that strong performance requires explicit accommodation of severe class imbalance, anatomical symmetry, heterogeneous structure scales, and CBCT-specific artifact patterns, while its neighboring benchmarks indicate that future development is likely to move further toward semi-supervised learning, finer internal dental anatomy, and broader clinical workflow integration (LaBella et al., 18 Aug 2025, Tan et al., 15 Sep 2025, Wang et al., 2 Dec 2025).