The drill breaks through the last thin layer of softened dentin, and there it is: living pulp, glistening at the floor of a deep cavity. The dentist now has seconds to make a call that decides the tooth’s fate — cap this exposed nerve and hope it heals, or take the whole pulp out and start a root canal. It is one of the most experience-dependent judgments in dentistry. A new AI pulpotomy tool wants to help make it.
The 30-Second Version
- Researchers trained a self-supervised AI (a DINOv2 vision transformer) to look at a microscope image of an exposed tooth pulp and predict whether it is healthy enough to “cap” during a pulpotomy.
- It matched expert endodontists — no statistically significant difference — and clearly beat dental novices, hitting 95.7% accuracy on its internal test set.
- Its sharpest edge was caution: novices said “yes, cap it” far too often (specificity as low as 37.5%), while the AI kept its false approvals in check.
- The honest caveat: it is a proof of concept trained on 443 images from a single center, and its outside-world test used just 21 images — where accuracy slipped to 85.7%.
It sounds like the kind of split-second expertise that takes a decade of chairside practice to build. But a 2026 study in the Journal of Dentistry by Qianli Zhang, Xiaoyan Wang and colleagues asked a pointed question: could a machine learn that judgment from pictures alone? Their tool — a self-supervised vision transformer trained on 443 microscopic pulp-stump images — is promising enough to take seriously and limited enough to keep firmly in check.
The study, in one glance
A pulpotomy is a tooth-saving compromise. When decay reaches the pulp, instead of removing the entire nerve, the dentist amputates only the inflamed top portion and seals the healthy stump underneath with a bioceramic “cap,” betting the pulp will survive. The whole gamble hinges on one visual judgment made under a microscope mid-procedure: is this exposed stump suitable to cap, or too damaged to trust? The team built an AI to classify that stump image as “suitable” or “unsuitable,” trained it on 443 high-resolution images, then tested it on an internal set of 93 images and an external set of 21 images pulled from the published literature. On both, they pitted the model against expert endodontists and against novice general practitioners.
What AI pulpotomy actually gets right
The impressive part is not that the AI scored high — plenty of models do that on their own test data. It is how it was right. The model reached 93.98% in cross-validation and significantly outperformed the classical supervised convolutional neural networks the team compared it against. On the internal test set it hit 95.70%, statistically comparable to expert endodontists (p > 0.05). For this narrow task, a machine approximated a specialist’s eye on a genuinely hard problem: “is this pulp healthy enough?” has no crisp line the way a fracture or a cavity does.
But does 95.7% survive contact with the real world?
Here is where a careful reader should slow down. That headline figure came from the internal test set — images drawn from the same clinical pipeline the model learned on. When the researchers turned it loose on 21 external images sourced from the published literature, accuracy fell to 85.71%. That is still respectable, but a roughly ten-point drop on unfamiliar data is exactly the pattern that separates a lab demo from a clinical tool. And 21 images is a very thin slice of the world’s pulps.
⚠ A proof of concept, not a product
The authors themselves frame this as a proof-of-concept tool. It was trained on 443 images from one institution and validated externally on just 21. Nothing here tells us how it behaves across different microscopes, lighting, cameras, patient populations, or the messy, blood-flecked reality of a live cavity.
Think of it less as a verdict and more as a very well-read second opinion — one that has studied thousands of textbook pulps but has never once watched a tooth heal over the following year.
That last point is the deepest limitation, and it is easy to miss. The AI was scored on whether it agreed with an expert’s visual classification, not on whether the capped teeth actually survived months later. Pulpotomy success is ultimately a biological outcome — does the pulp stay alive and symptom-free? This study measures agreement with a snapshot judgment, not healing. A tool can match the eye of an endodontist and still be wrong about the tooth, because even endodontists’ eyes are an imperfect proxy for what the tissue will do.
If the AI says “cap it,” who owns the outcome?
Decision-support tools have a way of quietly becoming decision-makers. If a model tells a busy or junior dentist that a marginal pulp is “suitable,” the tooth is capped, and it fails a year later — where does responsibility sit? The safe framing is that the AI flags and suggests while the licensed clinician decides and is accountable. That line is easy to write and hard to hold once a tool is fast, confident, and usually right.
Who actually gets this in the chair?
There is an access twist buried in the method. The AI reads microscope images of the pulp — which means it is most naturally deployed exactly where an operating microscope and digital capture already exist: well-resourced, specialist-heavy practices. Yet the clinicians who would benefit most from a “should I cap this?” safety net are early-career and general dentists in lower-resource settings, who are also the least likely to have the microscope rig the tool depends on. Without deliberate effort, a technology built to close the experience gap could end up widening the equipment gap instead.
What this means for you
If you’re a patient
Your dentist is not using this tool yet — it is early research. But it is a hopeful sign that the hardest, most subjective calls in tooth-saving dentistry are getting objective backup. If you are ever told a deeply decayed tooth might be saved with a pulpotomy rather than a root canal, that recommendation still rests on your dentist’s trained judgment, not an app.
If you’re a clinician
The most interesting result is not the accuracy — it is the specificity. Novices in this study over-diagnosed suitability (specificity as low as 37.5%), effectively saying “cap it” when they shouldn’t. A tool that reins in that optimism could be a genuine teaching aid. Just treat its “suitable” as a prompt to look harder, never as permission.
The bottom line
This AI learned the harder half of clinical wisdom: not the enthusiasm to save a tooth, but the restraint to know when a nerve is too far gone. That is real, and it is useful. But it read pictures, not patients — it never watched a single capped tooth heal or fail. Treat it as a sharp-eyed assistant that has read every textbook and seen no follow-ups, not an oracle that knows how the story ends.
Frequently asked questions
What is a pulpotomy, and why is the “cap it or not” decision so hard?
A pulpotomy removes only the inflamed upper part of a tooth’s pulp and seals the healthy remainder with a protective bioceramic cap, aiming to keep the tooth alive instead of doing a full root canal. The hard part is judging, by eye and mid-procedure, whether the exposed pulp is healthy enough to heal under that cap — a subjective call that depends heavily on the clinician’s experience.
Does 95.7% accuracy mean my tooth has a 95.7% chance of surviving?
No. That number is how often the AI’s classification of a pulp image agreed with expert judgment on the internal test set. It measures agreement with a specialist’s visual assessment, not the biological outcome of whether the capped tooth actually stays alive and symptom-free over time.
Is this AI available at my dentist’s office?
Not yet. The authors describe it as a proof-of-concept tool. It was trained on 443 images from a single center and tested externally on only 21 images, so it needs much broader validation before any clinical use.
Why did dental novices do worse, and does that make the AI “safer”?
Novices tended to over-approve capping — high sensitivity but low specificity, as low as 37.5% — meaning they too often judged a questionable pulp as suitable. The AI kept a more balanced, cautious profile with high specificity, comparable to expert endodontists. In this narrow test that restraint is an advantage, but it is not a substitute for a trained clinician’s overall judgment.
Could a tool like this eventually replace the dentist’s judgment?
The evidence doesn’t support that. It handles one isolated visual decision, was validated on a tiny external set, and was never tested against real healing outcomes. The realistic role is decision support — a second opinion the clinician weighs, with accountability staying firmly with the human.
Source & author credit
This article interprets, and does not reproduce, the following peer-reviewed study. All figures are the authors’ original findings.
Zhang Q, Hu M, Peng J, Wang X. Deep learning-driven intraoperative assessment of pulp stumps for precision pulpotomy. Journal of Dentistry. 2026;173:106796. DOI: 10.1016/j.jdent.2026.106796
ORCID — Qianli Zhang 0000-0003-1809-0748; Meiyu Hu 0009-0005-7355-7402; Xiaoyan Wang 0000-0002-8763-289X.
© 2026 Elsevier Ltd. All rights reserved; this is a proprietary (non-open-access) article, summarized here under fair-use commentary with a link to the original. Related Decadentry reading: AI vs. Endodontist: Who Reads a Dental X-Ray Better? Editorial standards: About the Author & Editorial Standards. Decadentry is an independent educational publication and is not affiliated with the study’s authors.
Leave a Reply