Skip to content
Endodontics & Restorative

AI Restorative Planning: A Blueprint or a Blur?

AI restorative planning topped 99% at spotting a post and ferrule, yet its tooth outlines matched experts’ by only about half.

Dentist in protective gear performing a dental procedure on a patient in a modern clinic setting.

A molar has had its root canal. The nerve is gone, the tooth is brittle, and before a crown can go on, someone has to answer two unglamorous questions: is there enough solid tooth left above the gum to grip a crown, and how long and wide should the metal post be that anchors it? Get either wrong and the restoration loosens, leaks, or splits the root. Those judgement calls usually need a prosthodontist and a careful look at an X-ray. A new study puts AI restorative planning to the test: can a pair of neural networks make them instead?

The 30-Second Version

  • Researchers built an AI restorative planning system that reads a dental radiograph and does two jobs: flag whether a tooth needs crown lengthening, and recommend post dimensions for a root-canal-treated tooth.
  • It topped 99% accuracy at the yes/no parts — spotting a post and judging whether there was enough ferrule.
  • But its tooth outlines overlapped the experts’ only about halfway: a mean IoU of 0.52 and a Dice score of 0.66.
  • The honest caveat: it was built at a single university, the abstract reports no external validation and no real patient outcomes, and sample sizes aren’t given. This is a prototype, not a product.

It sounds like the kind of quiet, practical AI dentistry actually needs — not a robot surgeon, just a second pair of eyes on a radiograph. The study, by Mohamed Mostafa Rezk, Lamiaa Said ElFadaly and Zeinab Abd Elhalim Taha at October University for Modern Sciences & Arts in Cairo, appears in The Journal of Prosthetic Dentistry (2026). The team trained two deep-learning pipelines to automate the radiographic groundwork of restoring an endodontically treated tooth, then measured how closely the machine’s reading matched the dentists’. The guiding question is the one that matters in the chair: can software do the measuring, or only the spotting?


The study, in one glance

The system splits the problem in two. A crown-measurement pipeline first uses a U-Net to segment each tooth on panoramic and periapical images, crops the individual tooth, then hands it to a Mask R-CNN that detects the tooth and calculates crown length — the numbers behind supracrestal tissue attachment and ferrule height, which decide whether a tooth needs crown lengthening before it can be restored. A separate post-measurement pipeline runs Mask R-CNN straight on the image to find existing endodontic posts and measure their length and width against established prosthodontic criteria. Annotations were drawn under the supervision of dental professionals, and the models were judged against those expert annotations.

>99%
accuracy identifying a post and assessing ferrule presence
0.52
mean IoU — how tightly its tooth outlines hugged the experts’
0.66
Dice coefficient — the companion overlap score for the segmentation
High at spotting, fuzzy at measuringSure about yes/no, fuzzy on the rulerScores rescaled to 0–100 for comparisonSpot a post& ferrule>99%Outline overlap(Dice)0.66Outline overlap(IoU)0.52A yes/no judgement can be near-perfect while the pixel-by-pixel measurement stays only middling.
The gap between spotting something and measuring it precisely. Figure: Decadentry, based on data reported in the study (DOI: 10.1016/j.prosdent.2026.10.009).

Where AI restorative planning earns its keep

The headline number is real and worth respecting. Deciding whether a radiograph shows a post, and whether a tooth has enough ferrule — that collar of sound tooth a crown needs to grip — is exactly the kind of structured yes/no task deep learning is good at, and the system cleared 99%. The authors report that it flagged supracrestal tissue attachment violations and produced post-size recommendations that lined up with expert-guided decisions. For a general dentist in a clinic with no prosthodontist down the hall, a tool that reliably says “this tooth doesn’t have enough ferrule — don’t just cement a crown and hope” is not nothing. It is the sort of quiet checklist automation that catches the error before it becomes a remake. The team also notes a GPU cut training time by more than 70%, which is an engineering convenience rather than a clinical finding, but it tells you this could run on ordinary research hardware.

So why does the ruler blur?

Because spotting and measuring are different skills, and the study is honest enough to report both. The segmentation model — the part that draws the outline of each tooth before anything gets measured — scored a mean IoU of 0.52 and a Dice coefficient of 0.66. Those are overlap scores: how much the machine’s traced tooth shares with the expert’s traced tooth. An IoU of 0.52 means the two outlines agree on roughly half their combined area. For telling a post from no post, that is plenty. For deciding that a ferrule is 1.8 mm rather than 2.2 mm, or that a post should be 8 mm not 9 mm, half-right outlines are a shaky foundation, because every millimetre of that measurement rides on where the edge was drawn.

⚠ A high score and a soft score, side by side

The same system reports “>99%” and “0.52” about itself. They aren’t in conflict — they measure different things. The 99% is a decision; the 0.52 is a measurement. Read the headline number without the overlap number and you will badly overestimate what the tool can do with a dimension that matters to a tenth of a millimetre.

It is a confident spotter and a hesitant surveyor — sure there is a fence, less sure exactly where it runs.

There is a second limit the abstract makes plain by omission. It reports no external validation — no test on radiographs from other clinics and other X-ray machines — and no clinical outcomes, only agreement with the annotations it was trained to match. It also doesn’t state how many teeth or images were used. A model that matches its own graders on its own data has shown it can learn the task, not that it will hold up on the messier image off a different sensor in a different surgery.

Who is accountable when the post is wrong?

This is where “decision support” earns its keep. If the software recommends a post dimension and the dentist follows it and the root fractures two years later, the responsibility still sits with the clinician who placed it — as it should. That is a feature, not a loophole: the system is built to supply a measurement and a flag, and the human is meant to weigh them against the tooth in front of them. The danger is the opposite of distrust. A number on a screen carries false authority, and “the AI said 8 mm” is an easy thing to lean on when you’re busy. A 0.52 IoU is a standing reminder that the number deserves a second look, not a signature.

Does it reach the clinics that need it most?

The authors pitch this squarely at resource-limited settings, where a prosthodontist’s eye is scarce, and that framing is fair — it is the most useful version of this idea. But it is also where an unvalidated tool is riskiest, because there is less specialist oversight to catch a bad recommendation. Equity here means validating the model on exactly those clinics’ radiographs — their scanners, their patient mix, their image quality — before it is trusted to stand in for the expert who isn’t there. A tool justified by scarcity has to be proven under scarcity, not only in the lab that built it.

What this means for you

If you’re a patient

Nothing about your next crown changes. This is early research software, not something in your dentist’s operatory. If a root-canal-treated tooth of yours is being restored, the questions that matter — enough tooth left to hold a crown, the right post — are still answered by your dentist, possibly with imaging help, not by an app.

If you’re a clinician

Treat the two numbers separately. A near-perfect spot-the-post model and a 0.52-IoU measurement model can live in one paper. For anything dimensional — ferrule height, post length — verify against your own read of the radiograph. Watch for external validation on non-source data before letting a tool like this influence a restorative plan.

The bottom line

This is a useful, honest little study about a genuinely helpful task: flagging the teeth that need crown lengthening and roughing out a post. It is very good at the yes/no and only middling at the millimetre, and it says so plainly. The right way to read it is the way the authors seem to intend — an assistant that points and suggests, checked by the clinician who measures and decides. A spotter, not a surveyor; an aide, not an oracle.

Frequently asked questions

What does this AI restorative planning system actually do?

It reads a dental radiograph and performs two tasks for a root-canal-treated tooth: it judges whether the tooth needs crown lengthening — by assessing ferrule height and supracrestal tissue attachment — and it detects and measures an endodontic post, recommending dimensions against standard prosthodontic criteria. It uses a U-Net to outline teeth and Mask R-CNN models to detect and measure.

Is it accurate enough to use on patients?

Not yet. It topped 99% on the yes/no tasks (is there a post, is there enough ferrule), but its tooth-outlining overlapped experts’ by only about half — a mean IoU of 0.52 and a Dice coefficient of 0.66 — which is a weak base for precise measurement. The abstract reports no external validation and no clinical outcomes, so it is a research prototype, not a clinical tool.

Why report both “over 99%” and “0.52” — isn’t that contradictory?

No. The 99% is how often the system correctly made a decision — spotting a post, flagging ferrule presence. The IoU of 0.52 and Dice of 0.66 measure how precisely its traced tooth outlines matched the experts’, pixel by pixel. A model can be excellent at a yes/no call while only middling at an exact measurement; the two numbers describe different skills.

Could this help dentists without a prosthodontist nearby?

That is the authors’ stated goal, and the promise is real — a tool that flags “not enough ferrule” could prevent failed restorations in clinics with limited specialist access. But resource-limited settings are also where an unvalidated model is riskiest, so it needs to be tested on those clinics’ own radiographs and machines before being relied on.

Does it replace the prosthodontist?

No. It is decision support: it supplies a flag and a measurement, and the clinician weighs them against the actual tooth and makes the call. Responsibility for the restoration stays with the dentist, and the modest overlap scores are a reminder to verify any dimension the software suggests.

“It is a confident spotter and a hesitant surveyor — near-perfect at telling a post from no post, only half-sure exactly where the tooth’s edge runs.”

Source & author credit

This article interprets, and does not reproduce, the following peer-reviewed study. All figures are the authors’ original findings, as reported in the published abstract.

Rezk MM, ElFadaly LS, Taha ZAE. Dual-phase deep learning recommendation system for radiographic crown lengthening assessment and post selection for the restoration of endodontically treated teeth. The Journal of Prosthetic Dentistry. 2026; advance online publication. DOI: 10.1016/j.prosdent.2026.10.009

ORCID — none listed in the Crossref record for this article at the time of writing.

© 2026 The authors / Elsevier. This is a copyrighted article (not open-access); Decadentry summarises and interprets it under fair use and links to the original via its DOI. Figures here are Decadentry’s own, based on reported data. Verification level: abstract-verified (no open full text available). Decadentry is an independent educational publication and is not affiliated with the study’s authors. See our editorial standards and related reading on AI crown design and the full endodontics & restorative hub.

HB

Hossein Boustani Hezarani

Dentist · AI-in-Healthcare researcher · Founder of Decadentry

Hossein writes Decadentry to translate peer-reviewed dental research into clear, honest, jargon-free reading — celebrating what AI can do for dentistry while asking the hard questions the hype skips. Every article is checked against its primary source.

Share

Decadentry explains published research for education. It is not medical or dental advice — talk to a qualified clinician about your own care. Read our medical disclaimer and editorial standards.

Leave a Reply

Your email address will not be published. Required fields are marked *