Can AI Approve an Implant Plan?
Updated: Sep 24
Clinically reviewed by Jacob Shrayman, DDS, a general dentist who places implants using TrueGuide surgical guides in his own practice.
Short answer: no. Software can propose an implant plan. It cannot approve one, and no laboratory should ship a guide on the strength of a plan nothing human has looked at.
That is not a philosophical position about machines. It is a description of where the current generation of planning software actually fails, and of what those failures look like on screen — which is the part that matters, because they do not look like failures. They look like finished plans.
This article covers what the automation in implant planning genuinely does well, the specific places it breaks, why a broken result is hard to see, and what a plan review is for. It also gives you the three checks worth making before you approve any plan a laboratory sends you, ours included.
What the software is actually doing
When a planning platform reads a CBCT, several different things are happening and they are not equally reliable. It helps to separate them.
Segmentation is the step that turns a stack of gray-value slices into named objects: this volume of voxels is bone, that one is a tooth, that tube is the inferior alveolar canal. Modern platforms do this with trained models rather than pure thresholding, and on a clean scan of a dentate patient the result is good.
Landmark and canal tracing proposes the path of the mandibular canal, the sinus floor, the nasal floor. This is the output most often shown to the doctor as a finished thing, drawn as a smooth tube through the mandible.
Registration aligns the CBCT with the intraoral scan so the plan can be made in one coordinate space. This is the step that most often goes wrong quietly, and we come back to it below.
Position suggestion proposes an implant position from the prosthetic plan and the available bone — crown-down, centered in the ridge, a set distance off the canal.
All four are useful. None of them is a decision. Segmentation says where the model thinks the bone edge is. It does not say whether that edge is real, whether the site is restorable, whether the patient can tolerate the procedure, or whether the plan you are looking at belongs to the patient in your chair.
Where segmentation breaks
The failure modes are not random. They cluster in exactly the cases that are hardest clinically, which is the uncomfortable part.
Metal. Existing crowns, posts, amalgam, orthodontic retainers and previously placed implants scatter the beam. Around a metal object the reconstruction produces dark streaks and bright bands that are not tissue, and a segmentation model has no way to know they are not tissue. The segmented bone surface near a restored neighboring tooth can be drawn several tenths of a millimeter off, or can pick up a streak as if it were cortical plate. The arches that most need a guide are frequently the arches with the most metal in them.
Thin buccal plate. A buccal plate thinner than roughly the voxel size does not reliably appear in the volume at all. It is not that the model misreads it — it is that there is nothing in the data to read. The segmentation then shows a clean ridge with a continuous surface and no warning, and the plan is made against a contour that is missing its most fragile wall. This is a partial-volume limitation of the acquisition, not a software defect, and no amount of model training fixes it.
Low-contrast boundaries. In soft bone, where the cortical outline fades into cancellous, the edge the model draws is an estimate. Two runs of the same software on the same scan can put that edge in slightly different places.
The canal where it is hardest to see. Automatic canal tracing is strongest in the body of the mandible, where the canal is a well-defined dark tube in dense bone. It is weakest anteriorly, near the mental foramen and the anterior loop, and in patients where the canal wall is not corticated — which is common, and more common with age. The traced path is drawn smooth and continuous regardless. A smooth line through a region where the canal is genuinely ambiguous in the data is a confident-looking answer to a question the data cannot answer.
Anatomical variants. Bifid canals, accessory foramina, sinus septa, retromolar canals. Trained models are trained on what is typical. The variant is by definition what the training data has least of, and the variant is the case where a missed structure hurts.
The errors do not announce themselves
This is the whole point, and it is worth stating plainly.
When a segmentation goes wrong, the software does not return an error. It returns a model. The model has smooth surfaces, plausible anatomy and a clean rendering, because smoothing and surface reconstruction are applied to every output whether the underlying voxels supported it or not. A canal traced through noise looks exactly like a canal traced through a clearly corticated tube. A ridge contour built from scatter looks exactly like a ridge contour built from bone.
Published work on deep-learning segmentation of CBCT reports strong overall agreement with expert tracing, and that is a real result. But an average agreement figure across a dataset tells you about the dataset. It does not tell you about the case in front of you, and the cases that pull the tail of that distribution are the metal-heavy, thin-plate, low-contrast, atypical ones. When you are planning a single case, you are not experiencing the mean. You are experiencing one draw, and you have no way to know from the rendering which part of the distribution you drew from.
That is the argument for a person looking at the slices — not at the three-dimensional model, at the slices — before anything is manufactured. A human reviewer looking at axial and cross-sectional views can see the streak, can see that the plate disappears, can see that the canal outline is genuinely indistinct in this region and that the traced line is a guess. The reviewer can then plan with a margin appropriate to a guess, or ask for a different record.
Registration is the quiet one
Everything above concerns one file. The step that invalidates a plan most completely, and shows it least, is aligning two files.
The implant plan is made in the CBCT. The guide is manufactured against the intraoral scan, because the guide sits on teeth or on tissue and the scan is what describes those surfaces accurately. If the alignment between the two is off by half a millimeter, every trajectory in the plan is off by that amount at the tissue and by considerably more at the apex — and the plan will render beautifully, because the plan is internally consistent. The error lives between the two datasets, not inside either one.
Automatic registration algorithms are good and they still fail on predictable inputs: scans with few clear registration features, arches with heavy scatter obscuring the crown surfaces the algorithm wants to match on, edentulous arches with no rigid landmarks at all, and cases where the intraoral scan and the CBCT were taken far enough apart in time that the mouth has changed.
Validating registration is not something you can do from a rendered model. It is done by stepping through cross-sections and checking that the scanned crown surface sits on the radiographic crown, at several separated points around the arch — not one. If they diverge anywhere, the alignment is wrong everywhere, and no downstream accuracy figure means anything.
We check this on every case before a plan goes out, and it is one of the more common reasons a case comes back to a doctor for a new record.
What a plan review is for
TrueGuide plans are produced by a technician and reviewed with the doctor before anything is manufactured. That review is not a formality and it is not a sales call. It exists because of everything above.
What gets looked at:
The cross-sections at each planned site, in the slices, not the rendering — is the bone that the plan is relying on actually in the data, or is it a reconstruction of a region the scan could not resolve?
The canal trace against the raw volume, especially anteriorly.
Registration, checked at several separated points around the arch.
Whether the prosthetic plan the implant position serves is one the doctor actually intends — the restoration drives the position, and the restoration is a clinical decision.
Whether the guide design is seatable: support, access, irrigation, whether the patient can open far enough for the drill and the handle at that site.
The last one is a good example of something no planning software evaluates, because it is not in the files. Interocclusal opening at a second molar is not a property of a CBCT.
Three checks to make before you approve any plan
These are cheap, they take under a minute, and they catch the class of error that accuracy statistics never touch. Make them on plans from any laboratory.
1. Patient identity. Confirm the patient name and date on the plan against the chart, and confirm that the anatomy on screen is consistent with the patient you examined. Files get mismatched in every workflow that moves files, and a plan built on the wrong scan is internally perfect.
2. Tooth numbering and site. Confirm the site by number and by looking at it, not by the label alone. Numbering conventions differ between software packages and between countries, and a plan can be transcribed one position off. Count from the midline on the screen.
3. Plan version. Confirm that what you are approving is the latest version and that the version number or date on the file matches the one from the review call. When a plan is revised, the superseded file does not disappear from anyone's downloads folder. Approve a version, in writing, and have the guide manufactured against that version.
We put the version and date on every plan we send for exactly this reason.
What automation is genuinely good for
None of this is an argument for doing the work by hand. Automation earns its place in this workflow, and it earns it in the same places most automation does — repetitive, well-defined, verifiable work.
A proposed canal trace that a technician then corrects is faster and, in our experience, better than one drawn from scratch, because the technician spends their attention on the ambiguous regions rather than on the obvious ones. A first-pass segmentation is a good starting surface. A crown-down position suggestion is a sensible default to argue with. Automated checking of mechanical constraints — sleeve-to-bone distance, drill length against available height, minimum wall thickness in the guide body — is work software does more reliably than a person, because it is arithmetic, and it never gets tired at the end of a long day.
The distinction that matters is between software that proposes and software that decides. Proposal plus review is faster and safer than either alone. Proposal treated as decision is where the trouble lives.
Who is responsible for the plan
Worth being clear about, because the phrase AI-planned sometimes gets used as though it moved responsibility somewhere.
It does not. In the United States a patient-specific surgical guide is regulated as an endosseous dental implant accessory under 21 CFR 872.3980, product code NDP, which is Class One and exempt from premarket notification. The treating dentist prescribes the treatment, approves the plan and places the implant. A laboratory produces a device to that approved plan. Planning software is a tool the laboratory uses, in the same category as a scanner or a printer — the technician is accountable for its output, and the doctor is accountable for the treatment.
No planning platform on the market today is cleared to make an autonomous treatment decision, and none of them claim to be. Where you see automation described as approving a plan, that is marketing language describing a workflow in which a person approved it.
What to ask a laboratory that says its planning is AI-driven
Four questions, and the answers are informative.
Who looks at my case, and at what stage? A name and a step, not a statement that the system checks it.
How is the CBCT-to-scan registration validated? If the answer is that the software handles it, the answer is that it is not validated.
What happens when a record is not good enough to plan from? A laboratory that has never sent a case back is either seeing extraordinary records or is planning from records it should not be planning from.
Do I see the plan before it is manufactured, and can I change it? The answer should be yes, in cross-section, with the version recorded.
Bottom line
AI in implant planning is real, it is useful, and it is not a reviewer. Its characteristic failure is a confident, smooth, entirely plausible model of anatomy that is not there — produced most often in the metal-heavy, thin-plate, atypical arches that most need a guide in the first place. Nothing in the rendering tells you which case you have.
A person looking at the slices catches it. That is the whole reason the review step exists, and it is why TrueGuide will not ship a guide against a plan no one has reviewed with the treating doctor.
If you want the record specification that keeps cases out of the ambiguous territory in the first place, that is covered in CBCT and scan records for guided surgery.
This article is about laboratory workflow and planning software. It is not a treatment plan for any patient, and it is not medical advice. Treatment decisions belong to the treating doctor.
Comments