AlphaDent: Dental Benchmark & Image Dataset
- AlphaDent is a dental image dataset characterized by high-resolution DSLR intraoral photographs and detailed YOLO-compatible mask annotations for pathology segmentation.
- It provides a 9-class taxonomy for instance segmentation, supporting applications like object detection, pathology localization, and multi-task analyses.
- The benchmark demonstrates differences in performance between fine-grained caries detection and broader pathology classification, offering reproducible evaluation protocols.
Searching arXiv for the AlphaDent paper and closely related dental imaging work to ground the article with fresh citations. AlphaDent is an openly released dental image dataset and benchmark for automated tooth pathology detection from high-resolution DSLR intraoral photographs, with a primary emphasis on instance segmentation. It was introduced as a response to the sparse public ecosystem of dental imaging datasets, which the authors describe as heavily skewed toward radiographs and often lacking clear licenses, standardized annotations, or pathology-focused support for photographic intraoral imagery. In that framing, AlphaDent is not only a dataset release but also a reproducible benchmark consisting of open images, YOLO-compatible labels, open training and inference code, pretrained weights, and a hidden-label evaluation setup (Sosnin et al., 30 Jul 2025).
1. Dataset identity and intended use
AlphaDent is “primarily intended to solve the Instance Segmentation problem,” and each pathology instance is annotated not only with bounding boxes but also with “carefully marked masks” (Sosnin et al., 30 Jul 2025). The target objects are visible dental pathologies and restorations or conditions observable in intraoral photographs, rather than radiographic findings or whole-tooth anatomical segmentation. Because the dataset provides class-specific masks for distinct tooth defects and restorative findings, it can also support object detection, pathology localization, per-class prevalence analysis, and potentially weakly supervised or multitask learning.
The authors position AlphaDent as a first fully open dataset of DSLR-based intraoral dental photographs with mask annotations for multiple pathology classes. Its distinguishing contribution is therefore not sheer size, but openness, modality, and task focus. Relative to prior public datasets described in the paper, AlphaDent fills an open-data gap for high-resolution photographic pathology segmentation: earlier public resources are described as stronger in radiography, tooth segmentation, cephalometric detection, CBCT-heavy multimodal data, or 3D intraoral scans, whereas AlphaDent concentrates on pathology-aware instance segmentation in intraoral photographs (Sosnin et al., 30 Jul 2025).
A later benchmark paper, “Caries DETR: Tooth Structure-aware Prior and Lesion-aware Dynamic Loss Refinement for DETR Based Caries Detection,” treated AlphaDent as a public benchmark for object detection in intraoral images, which indicates that the dataset quickly became a methodological reference point beyond its original YOLO segmentation baseline (Liu et al., 26 Apr 2026).
2. Image acquisition, composition, and clinical content
AlphaDent consists of anonymized photographs from 295 real patients, totaling 1320 photographs according to the Methods section and “over 1200 images” in the abstract and conclusion (Sosnin et al., 30 Jul 2025). The images were collected from January 2024 to January 2025. Several images are available per patient, with an average of 4.5 images per patient. The acquisition setup used an intraoral mirror, a Canon 6D Mark II camera, and a Canon 100mm f/2.8L Macro IS USM lens.
The image set includes both frontal intraoral images and separate occlusal-view images of the upper and lower jaws. Most images have a resolution greater than 5000 by 3000 pixels, approximately 15 megapixels. This high-resolution photographic regime distinguishes AlphaDent from smartphone-only oral RGB screening studies and from radiograph-dominant dental datasets; a related smartphone study, for example, addressed binary calculus classification on a dataset of 220 RGB images rather than pathology-aware instance segmentation on high-resolution DSLR photographs (Garg et al., 2023).
The paper encodes sex and age metadata in filenames and reports 131 male patients and 162 female patients in its statistics, although those counts sum to 293 rather than 295. The manuscript includes an age distribution plot, but no exact age summary statistics beyond per-image filename encoding. Clinically, visible problem areas were assessed from the photographs, and hidden carious lesions were inferred from local visual signs such as gray enamel and localized darkening suggestive of caries. Crowns included mainly metal-ceramic, metal, ceramic, and zirconium restorations. The abrasion class is grounded in detailed assessment of wear facets, including both pathological and physiological wear (Sosnin et al., 30 Jul 2025).
3. Annotation taxonomy and label format
The central annotation contribution is a 9-class taxonomy for pathology and restoration instance segmentation. The six caries subclasses follow the standard 5 Black classes plus a 6th class proposed by WHO, localized in “immune zones” such as molar and premolar cusps and incisal edges (Sosnin et al., 30 Jul 2025).
| Class ID | Class name | Definition |
|---|---|---|
| 0 | Abrasion | Mechanical wear of hard tissues |
| 1 | Filling | Fillings of various types |
| 2 | Crown | Installed crown |
| 3 | Caries 1 class | Caries in fissures and blind pits |
| 4 | Caries 2 class | Caries of contact surfaces of molars and premolars |
| 5 | Caries 3 class | Caries of contact surfaces of incisors and canines without cutting-edge damage |
| 6 | Caries 4 class | Caries of contact surfaces of incisors and canines with cutting-edge damage |
| 7 | Caries 5 class | Cervical caries of vestibular and lingual surfaces |
| 8 | Caries 6 class | Caries of cutting edges of front teeth and cusps of chewing teeth |
Annotations are polygon-based. For each image there is a corresponding .txt file with one row per feature instance, using the exact YOLO segmentation line format:
<class-index> <x1> <y1> <x2> <y2> ... <xn> <yn>
The boundary coordinates are normalized to values from 0 to 1. In addition, each image has a folder of binary or instance masks, with filenames such as 01_class_5.png, where the prefix is the mask ID within the image and the suffix encodes the class ID from 0 to 8. The paper states that the labels directory contains “markup for images in the form of polygons and masks,” and that “in addition to the bounding boxes, carefully marked masks are available for each pathology” (Sosnin et al., 30 Jul 2025).
The masks were manually produced by dentists using CVAT. The manuscript, however, explicitly notes a quality-control limitation: the dataset was “mainly prepared by one specialist,” and markup by several specialists would increase annotation accuracy. No inter-annotator agreement study, multi-rater adjudication process, or formal annotation audit is reported.
4. Data splits, benchmark protocol, and reported performance
The labeled development data was divided into training and validation by patients rather than by photographs, which reduces patient-level leakage. Specifically, 273 patients were included in training and 22 patients in validation. Table 3 reports that validation covered 83 images containing 872 annotated instances across all 9 classes. The hidden-label test set was assembled later, in May 2025, and labeled separately to simulate deployment on new data. The paper states that the test data contains 135 images, although the directory-structure description lists the images/test folder as containing 121 images, which the manuscript leaves unresolved (Sosnin et al., 30 Jul 2025).
The benchmark experiment uses YOLOv8x for instance segmentation with 9 classes, image sizes of 640 px and 960 px, 100 epochs, and a “Large set of augmentations.” No optimizer, learning rate, batch size, scheduler, hardware details, or explicit objective-function formulas are reported. Evaluation uses precision, recall, [mAP](https://www.emergentmind.com/topics/mean-average-precision-map)@50, and mAP@50:95, although one table header prints mAP@50:90 while the accompanying description defines the metric as averaging IoU thresholds from 0.50 to 0.95. The paper itself indicates that this is likely a notation inconsistency.
At 960 px on the validation set, the 9-class benchmark reports an overall precision of 0.513, recall of 0.430, mAP@50 of 0.436, and mAP@50:90* of 0.240. Classwise performance is highly uneven: abrasion reaches mAP@50 = 0.768, crown 0.841, and filling 0.677, whereas Caries 2 reaches 0.253, Caries 3 reaches 0.052, Caries 4 reaches 0.293 on only 4 validation instances, and Caries 6 reaches 0.043 on only 5 validation instances (Sosnin et al., 30 Jul 2025).
The authors therefore also report a 4-class ablation in which all caries subclasses are merged into a single caries category.
| Setting | Classes | Overall mAP@50 |
|---|---|---|
| Fine-grained benchmark | 9 | 0.436 |
| Caries-collapsed benchmark | 4 | 0.680 |
In the 4-class setting, the overall validation metrics become precision 0.626, recall 0.690, mAP@50 0.680, and mAP@50:90* 0.458. Crown is the best-performing category with mAP@50 = 0.893, followed by abrasion at 0.808 and filling at 0.701, while the merged caries category remains difficult at 0.316. This suggests that AlphaDent is already useful for broader pathology detection and segmentation, but that fine-grained six-way caries categorization is limited by class imbalance and subtle inter-class visual differences (Sosnin et al., 30 Jul 2025).
The training distribution helps explain this behavior. The training data contains 5957 abrasion masks and 2187 filling masks, but only 43 masks for Caries 4 and 52 masks for Caries 6. The validation set is correspondingly sparse for those classes: 4 instances for Caries 4 and 5 for Caries 6. The paper explicitly notes that “there are some types of caries for which there is little data and which are poorly determined by the model, especially for the caries types 3, 4, and 6.”
5. AlphaDent as an open benchmark in later research
AlphaDent’s role expanded beyond its original release because it provides a hidden-label Kaggle evaluation with mAP@50 ranking, YOLO-ready labels, open code, and pretrained weights. Access is organized around a released dataset structure, an images/ and labels/ layout, a yolo_seg_train.yaml file describing classes, a public GitHub repository at https://github.com/ZFTurbo/AlphaDent, and a Kaggle competition at https://www.kaggle.com/competitions/alpha-dent/ (Sosnin et al., 30 Jul 2025).
Subsequent work treated AlphaDent as a nontrivial benchmark for caries detection. “Caries DETR” describes AlphaDent as containing 1,455 intraoral images with pixel-level labels for Abrasion, Filling, Crown, and Caries 1–6, and evaluates the dataset in an object-detection setting rather than instance segmentation. In that study, a DINO-based detector augmented with Tooth Structure-aware Query Initialization and Lesion-aware Dynamic Loss Refinement achieved mAP = 29.9, mAP50 = 50.5, and mAP75 = 29.6 on AlphaDent with a ResNet-50 backbone, outperforming several CNN and DETR-family baselines (Liu et al., 26 Apr 2026).
That later usage is significant for two reasons. First, it shows that AlphaDent can function as a benchmark outside the original YOLO segmentation formulation, including in COCO-style detection pipelines. Second, the later paper also confirms that the fine-grained caries categories remain the hardest part of the benchmark: under Caries-DETR, classwise AP on AlphaDent remains low for several subclasses, including Caries 2, Caries 3, and Caries 6, even when overall mAP improves (Liu et al., 26 Apr 2026). A plausible implication is that AlphaDent’s enduring research value lies less in easy common classes and more in its capacity to stress-test fine-grained lesion localization under subtle appearance variation.
6. Limitations, reporting inconsistencies, and research implications
The dataset paper explicitly acknowledges several limitations. Annotation was “mainly prepared by one specialist,” which raises concerns about subjectivity and label noise, particularly for visually inferred hidden caries. Most of the dataset was collected “on one camera,” which may constrain generalization across different optics, illumination profiles, white-balance settings, or smartphone capture conditions. The collection context also appears geographically and clinically narrow, so broader population-level generalization is not established (Sosnin et al., 30 Jul 2025).
Class imbalance is substantial. The rarest categories—especially Caries 4 and Caries 6—have so few examples that benchmark reliability is intrinsically fragile. The test split is image-based rather than patient-based, unlike the train and validation split, so patient-level independence cannot be verified from the paper if multiple test images belong to the same patient. The manuscript also contains several internal inconsistencies: sex counts sum to 293 rather than 295; test image counts appear as both 135 and 121; and the metric header mAP@50:90 conflicts with the textual definition mAP@50:95.
These limitations do not negate the dataset’s importance. Rather, they define its current epistemic status. AlphaDent is strongest as an open, reproducible benchmark for pathology-aware photographic dental instance segmentation, especially for common categories such as abrasion, fillings, and crowns. It is weaker as a definitive substrate for reliable six-way fine-grained caries categorization without additional data, multi-expert annotation, and broader acquisition diversity. This suggests a clear research agenda: enlarge rare caries subclasses, diversify sensors and clinical environments, add multi-rater validation, and preserve the benchmark’s open-code and hidden-label evaluation structure while improving its annotation rigor (Sosnin et al., 30 Jul 2025).
