In recent years, several algorithms have been developed for the segmentation of the Inferior Alveolar Canal (IAC) in Cone-Beam Computed Tomography (CBCT) scans. However, the availability of public datasets in this domain is limited, resulting in a lack of comparative evaluation studies on a common benchmark. To address this scientific gap and encourage deep learning research in the field, the ToothFairy challenge was organized within the MICCAI 2023 conference. In this context, a public dataset was released to also serve as a benchmark for future research. The dataset comprises 443 CBCT scans, with voxel-level annotations of the IAC available for 153 of them, making it the largest publicly available dataset of its kind. The participants of the challenge were tasked with developing an algorithm to accurately identify the IAC using the 2D and 3D-annotated scans. This paper presents the details of the challenge and the contributions made by the most promising methods proposed by the participants. It represents the first comprehensive comparative evaluation of IAC segmentation methods on a common benchmark dataset, providing insights into the current state-of-the-art algorithms and outlining future research directions. Furthermore, to ensure reproducibility and promote future developments, an open-source repository that collects the implementations of the best submissions was released.

Background: Osteoporosis, a systemic skeletal disorder, is expected to affect 60% of women over 50. While dual-energy X-ray absorptiometry (DXA) scans are the current gold standard for diagnosis, they are typically used only after fractures occur, highlighting the need for early detection tools. Initial studies have shown panoramic radiographs (PRs) to be a potential medium, but these have methodological flaws. This study aims to address these shortcomings by developing a robust AI application for accurate osteoporosis identification in PRs. Methods: A total of 348 PRs were used for development, 58 PRs for validation, and 51 PRs for hold-out testing. Initially, the YOLOv8 object detection model was employed to predict the regions of interest. Subsequently, the predicted regions of interest were extracted from the PRs and processed by the EfficientNet classification model. Results: The model for osteoporosis detection on a PR achieved an overall sensitivity of 0.83 and an F1-score of 0.53. The area under the curve (AUC) was 0.76. The lowest detection sensitivity was for the cropped angulus region (0.66), while the highest sensitivity was for the cropped mental foramen region (0.80). Conclusion: This research presents a proof-of-concept algorithm showing the potential of deep learning to identify osteoporosis in dental radiographs. Furthermore, our thorough evaluation of existing algorithms revealed that many optimistic outcomes lack credibility when subjected to rigorous methodological scrutiny.

Partial-arch intraoral scans (IOSs) are commonly used in clinical dentistry where high precision and reduced scanning areas are required. However, most existing tooth segmentation algorithms are developed only for full-arch IOSs and perform poorly when applied to partial-arch data. This study aimed to develop a fully automated deep learning (DL) model for tooth segmentation and labeling on both full- and partial-arch IOSs. We collected 600 IOSs (300 full-arch and 300 partial-arch) from a dental clinic. The proposed model was based on a two-stage DL model (ToothInstanceNet), and incorporated four enhancements: (1) artificial partial-arch IOSs, (2) DL-based alignment module, (3) FDI-aware postprocessing algorithm, and (4) real partial-arch IOSs. Model performance was evaluated via 5-fold cross-validation using F1-score, tooth Dice, tooth macro-F1, and macro-IoU. In addition, the model was evaluated with the public Teeth3DS dataset, and we analysed correlations between dental conditions and model errors. The model achieved an F1-score of 0.9908 and 0.9884; tooth Dice of 0.9819 and 0.9862; tooth macro-F1 of 0.9940 and 0.9786; and macro-IoU of 0.9403 and 0.9280 on full- and partial-arch IOSs, respectively. The model also demonstrated superior performance (score = 0.9870) in 3DTeethSeg challenge. Correlation analyses revealed that certain dental conditions, particularly residual roots, residual crowns, missing teeth, and partially erupted teeth, were significantly and positively associated with the model's errors. The current study proposes the first fully automated method for tooth segmentation and FDI labeling on both full- and partial-arch IOSs. The final model demonstrated high accuracy for both scan types, indicating its potential for integration into clinical dental work. This work could aid clinicians in the first step of tooth identification in digital dental workflows, and lays the groundwork for extending the automation of the downstream applications, such as diagnosis and monitoring on partial-arch IOSs.

Considering the importance of distinguishing between primary and permanent teeth in children with mixed dentition, this study aimed to develop and evaluate an automated method for segmenting and labelling primary and permanent teeth in digital impressions. 716 digital impressions from 351 patients with primary or mixed dentitions were collected from the Netherlands, Brazil, and the 3DTeethSeg22 challenge dataset. The scans were annotated with tooth segmentations and primary and permanent teeth FDI numbers. A deep learning model was applied that combined large-context predictions for tooth labelling with high-resolution predictions for tooth segmentation. Using the collected scans, the model was trained and evaluated with five-fold cross-validation for tooth detection (F1-score), tooth segmentation (Dice score), and tooth labelling (macro-F1). Additionally, the model was trained and evaluated using the train-test split of the 3DTeethSeg22 challenge dataset. The developed model achieved highly effective results for tooth detection (F1-score = 0.996), tooth segmentation (Dice = 0.969), and tooth labelling (macro-F1 = 0.989). Moreover, a digital impression was processed in under two seconds on average. Furthermore, the proposed method outperformed the top-ranked 3DTeethSeg22 challenge submission (score = 0.954 vs. 0.976) and was particularly effective for tooth labelling (tooth identification rate = 0.910 vs. 0.955). Failure cases revealed mistakes for unusual dental conditions or ambiguous tooth eruption patterns. A highly effective algorithm for tooth segmentation was developed to differentiate between primary and permanent teeth in digital impressions. This fast and accurate model can benefit dentists in documenting children's teeth during the mixed dentition stage. The algorithm provides an accurate and reliable tool for AI-assisted identification and numbering of primary and permanent teeth in digital impressions obtained from children with mixed dentition, thereby enhancing clinical workflow, improving treatment planning accuracy, and facilitating communication with patients and caregivers.

Bitewings are commonly used radiographs for visualizing teeth and various dental conditions. Manual labeling and diagnosis on bitewings for chart filing are time-consuming and prone to observer-dependent variations. This multi-center study proposes a deep learning (DL) approach to automate comprehensive chart filing of bitewings. A total of 1045 bitewings from Germany and The Netherlands were used for training and validation, and 216 from Slovakia for external testing. Annotations were performed by two dentists, one PhD researcher, and one caries expert. Hierarchical Mask DINO was developed for multi-class hierarchical end-to-end instance segmentation. Unmodified Mask DINO, SparseInst, and Mask R-CNN were used for comparison. Model performance was evaluated using F1-score, sensitivity, specificity, precision, mean average precision (mAP), and area under receiver operating characteristic curve (AUC). Mask DINO models exhibited high effectiveness for tooth segmentation and labeling, achieving precision, sensitivity, and F1-scores above 0.96. Hierarchical Mask DINO outperformed the other models in dental finding classification. F1-scores for implant, crown, pontic, filling, root canal treatment (RCT), caries lesion, and calculus deposit were 0.944, 0.918, 0.952, 0.956, 0.988, 0.749, and 0.758, respectively, with specificities all above 0.95. This study presented a DL-based approach for comprehensive assessment and diagnosis of bitewings, underlining its potential to enhance the efficiency and accuracy of chart filing in dental practice. The proposed model provided fully automated tooth segmentation and numbering, along with comprehensive segmentation of dental conditions. Dental professionals can benefit from this model for reducing manual workload and enhancing clinical diagnosis.

Accurate interpretation of cone-beam computed tomography (CBCT) scans is critical for oral diagnosis and treatment planning. Existing methods for automated tooth segmentation in CBCT face challenges, such as difficulties in generalizing across imaging artifacts and anatomical variations, as well as requiring manual revisions in many cases. To address these limitations, this study introduces ToothSeg, a fully automated approach for tooth instance segmentation and numbering in CBCT using deep learning and self-correction. ToothSeg combines semantic and instance segmentation into a unified method where their respective strengths are complemented. In particular, self-correction is employed when combining the segmentations, resolving merged or split teeth and determining the optimal sequence of tooth numbers for each dental arch. We conducted a comprehensive evaluation using a diverse in-house dataset (n = 1282, 25+ devices) and the publicly available ToothFairy2 challenge dataset (n = 480, 1 device), including an ablation study, a comparison to state-of-the-art methods, and an analysis of challenging cases. Compared to an optimized semantic segmentation model, including instance segmentation and self-correction consistently improved tooth segmentation (True Positive Dice: 93.6% to 94.3%) and tooth detection and numbering (multiclass instance F1: 94.2% to 95.5%). Furthermore, ToothSeg outperformed the other methods on both datasets (True Positive Dice: $\boldsymbol{\ge }$ +0.4%, multiclass instance F1: $\boldsymbol{\ge }$ +1.8%), particularly for challenging cases. This study provides a promising approach for automated tooth segmentation and numbering in CBCT, which is significant for reducing manual workload and supporting scalable, data-driven research in oral and craniofacial health. Code and models are publicly available at https://github.com/MIC-DKFZ/ToothSeg.