Where to Find Dental AI Datasets: Public vs. Custom-Collected Data
Anyone who has actually tried to build a dental AI model knows the real punchline pretty quickly: the model architecture is rarely what breaks things. The data is. I’ve spent time on both sides of this, reading radiographs in a clinic and later helping data teams figure out why a “great” model fell apart the moment it saw images from a different X-ray unit. Almost every time, the problem traced back to the dataset, not the code.
So let’s get into where dental AI datasets actually come from, what’s realistically available in public repositories right now, and where the public options run out of road faster than people expect.
Types of dental datasets (briefly)
Dentistry works with a narrower set of imaging modalities than general radiology, but each one comes with its own quirks:
- Bitewing and periapical X-rays, the everyday intraoral shots used for caries and bone level checks
- Panoramic radiographs (OPGs), the wide jaw view that most public datasets are built around
- Cone-beam CT (CBCT), volumetric 3D imaging used for implants, endodontics, and surgical planning
- Intraoral photos and scans, used for soft tissue, occlusion, and orthodontic tracking
- Cephalometric radiographs, mostly for orthodontic AI
None of these are interchangeable in how they’re acquired, how noisy they are, or how they get labeled. That matters more than most people assume when you’re trying to combine sources or expect a model to generalize across them.
Public datasets available today
Where can you actually find dental X-ray datasets? There are more options than there used to be, but the pool is still shallow. Panoramic radiographs dominate the public landscape. Collections like DENTEX and OdontoAI provide annotated OPGs for tooth detection, numbering, and pathology classification, and researchers have pooled the two together to get past the limits of either one alone. Public periapical and bitewing collections exist too, but they tend to be smaller, older, or narrower in what they label.
Are there public CBCT datasets? Yes, though they showed up later and are still fairly rare compared to 2D radiograph sets. Projects like CTooth and CTooth+ released annotated CBCT volumes specifically for tooth segmentation, and a few newer multimodal releases have started pairing CBCT with panoramic and periapical images from the same patients. A 2024 systematic review of publicly available dental imaging datasets found that most existing collections are still 2D, single modality, and modest in patient count. CBCT and multimodal datasets remain the exception rather than the norm.
What datasets exist for tooth segmentation? This is genuinely the best covered task in public dental AI data. Between OdontoAI, DENTEX, CTooth, CTooth+, and a handful of Kaggle and GitHub hosted OPG collections, tooth level segmentation and FDI numbering have more open resources than almost anything else in dental AI. Caries staging, periodontal bone loss measurement, and treatment outcome prediction are much thinner by comparison.
Are public datasets enough for commercial AI?
For a class project or an early proof of concept, often yes. For something you’re planning to put in front of the FDA, or in front of a paying clinic, usually not on their own. A few reasons keep coming up:
Volume and diversity. Public dental datasets typically run from a few hundred to a few thousand studies, small compared to the tens of thousands of labeled images that robust deep learning pipelines usually want, especially for rarer pathologies.
Device and population bias. Most public sets were collected at one or two institutions on a narrow set of scanner models, from a fairly homogeneous patient population. A model that performs beautifully on that data can fall apart on a different sensor or a different demographic mix, and you often don’t find out until it’s already deployed.
Annotation inconsistency. Bounding boxes, 2D contours, and full 3D segmentations all get called “annotated,” but they aren’t interchangeable, and mixing them carelessly introduces label noise that won’t show up until validation, if you’re lucky enough to catch it there at all.
Regulatory expectations. Teams pursuing FDA clearance need training and validation data that actually reflects their intended use population and device types, and a patchwork of public datasets rarely satisfies that cleanly. The FDA’s own list of authorized AI/ML enabled medical devices is worth a look if you want a sense of how few dental specific clearances exist relative to radiology as a whole. That gap says as much about data maturity as anything else.
None of this makes public data useless. It’s genuinely useful for benchmarking, pretraining, and early prototyping. It just tends to be a starting point rather than a finish line once you’re building toward an actual product, which is something you can see play out across real projects in our case studies, where public data alone rarely covered the full scope of what a client needed.
Questions of anonymization: how dental imaging can identify a patient
This part gets underestimated constantly, and dental imaging has a wrinkle that a lot of general medical imaging doesn’t. CBCT and cephalometric scans capture facial bone structure, and in some fields of view, soft tissue, in enough volumetric detail to reconstruct a recognizable face.
Under HIPAA, full face photographic images and “comparable” images count as direct identifiers. The tricky part is that a 3D rendered face pulled from a CT or CBCT volume can qualify as comparable even though nobody handed over an actual photograph. Research on head CT de-identification has shown that facial reconstructions from volumetric scans can match a reference photo with striking accuracy, which is exactly the kind of re-identification risk regulators worry about. GDPR goes further and treats facial geometry as biometric data outright, putting it under stricter rules than ordinary health records.
Stripping DICOM metadata is the easy part of anonymization. The harder part is pixel level work, removing or distorting facial soft tissue geometry in CBCT without wrecking the diagnostic value of the jaw, teeth, and sinuses the model actually needs to learn from.
When custom data collection becomes necessary
In practice, teams tend to reach for custom collection once one of these becomes true:
The target pathology, early caries, specific lesion types, rare anomalies, is underrepresented in every public set they can find.
They need a particular scanner, field of view, or acquisition protocol to match the intended clinical deployment.
They need volumes large enough, and diverse enough across age, geography, and device, to hold up under a regulatory submission.
They need annotation depth, 3D segmentation or multi label diagnosis, that public sets simply don’t offer for that use case.
How production-ready datasets get built
A production dataset is never just “more images.” It’s a full pipeline: sourcing studies from real clinical settings, getting consent under the applicable framework, anonymizing at both the metadata and pixel level, and running annotation through quality control that catches inconsistent labeling before it quietly poisons a training run. Teams building dental imaging datasets and annotation services for X-ray and CBCT workflows are, in effect, solving the exact gaps described above. They fill in the modality coverage, pathology diversity, and annotation depth that public datasets leave open, while keeping de-identification and compliance part of the process from the start rather than something added on at the end.
FAQs
Is dental imaging data considered PHI? Yes. Beyond the obvious identifiers like name, date of birth, and medical record number, facial bearing imaging such as CBCT can function as a direct identifier under HIPAA’s “comparable image” language, and as biometric data under GDPR. That means anonymization has to happen at the pixel level, not just the metadata level.
How much data do you actually need to train a dental AI model? It depends a lot on the task and the architecture, but as a rough signal, public tooth segmentation datasets that perform well in benchmarks tend to run in the thousands of annotated images. Narrower pathology detection tasks with class imbalance often need meaningfully more to generalize reliably across devices and patient populations.
Can synthetic data replace real clinical images for dental AI? Synthetic and generative approaches are an active area of research and can help pad out thin classes or reduce privacy exposure, but they haven’t replaced real, clinically sourced data for regulatory grade models. Most peer reviewed studies still treat synthetic data as a supplement, not a substitute.
What’s the difference between bounding boxes and segmentation for dental annotation? Bounding boxes are faster and cheaper, and fine for localization tasks like flagging a region with a likely finding. Segmentation, whether 2D contouring or full 3D volumetric labeling, takes longer but is what’s needed when a model has to understand exact anatomical boundaries, such as for treatment planning or precise pathology measurement.
Are public datasets biased toward certain populations or devices? Often, yes. Most were collected at one or a small number of institutions using a limited set of scanner models, which can skew a trained model’s performance once it’s deployed on a different device or population than the one it learned from.
Where can I find dental X-ray or CBCT datasets to start experimenting? Panoramic radiograph datasets like DENTEX and OdontoAI, along with CBCT segmentation datasets like CTooth and CTooth+, are reasonable places to start prototyping. A recent systematic review of publicly available dental imaging datasets is a useful map of what else is out there before deciding whether you need to go further.
