Why dental implantation modeling breaks down at the planning stage
Implantation planning breaks down not during surgery, but earlier — during the analysis of computed tomography images.
Metal crowns, pins, and fillings create artifacts that "smear" bone density on adjacent slices, and patient movement during scanning and slice thickness add to this.
If such an image is accepted as is, the model is built on distorted data, and printing accuracy can no longer compensate for this.
Inaccurate segmentation is the second point of failure. The density threshold confidently separates cortical bone, but "sticks together" on thin walls, defects, and in the projection of the maxillary sinus.
The boundaries of the mandibular canal blur, bone thickness is overestimated — and an implant that in reality lacks volume ends up in the plan.
The third point is the gap between plan and surgery. The surgical guide is printed based on the same segmentation, and errors accumulate: model construction, guide seating on teeth, sleeve position, and drilling itself.
A total of one to one and a half millimeters results in either redoing the surgical guide and rescheduling the appointment, or placement with a risk of nerve damage, sinus perforation, and bone grafting instead of simple implant placement.
For the clinic, this is not one error but a chain: an occupied chair, repeated anesthesia, a new treatment plan for the patient, and a shift of the entire schedule. We analyze how this is closed — from input image control to guide verification before surgery.
What AI planning provides: position accuracy and engraftment prognosis
Planning relies on data that the clinic already collects: a series of CT slices and an intraoral scan in DICOM format.
The neural network marks the jaw, mandibular nerve canal, and sinuses, builds a bone density map, and provides the surgeon with a three-dimensional scene where the anatomy is already highlighted.
Manual marking takes hours and depends on the doctor's attention; automatic takes minutes and is reproduced identically from patient to patient.
Positioning and fit accuracy
The model iterates through options for axis, depth, and implant size and shows how each fits into the available bone volume.
Position deviation goes within tenths of a millimeter, and the surgical guide is printed from the same scene — insertion follows the calculated direction, not "by eye." The fit of the crown and framework becomes a consequence of the plan, not adjustment in the laboratory, so returns from the laboratory are several times fewer.
Engraftment and risk prognosis
We compare bone density and cortical layer thickness with the expected tightening moment, patient data, and observation history: age, diabetes mellitus, smoking, previous interventions.
The output is the probability of osseointegration for each scenario and a list of factors that reduce it. Where risk is high, the plan changes in advance: a different size, bone grafting, a two-stage technique.
Preparation of the surgical plan is reduced from several days to one visit — the patient leaves with a decision, not with waiting. The drop in the number of redoings is visible both in the chair and in laboratory workload.
Approaches to implantation: guide, navigation, personal model
All three approaches are based on tomography data and a digital model of the jaw. The difference is how this model is transferred to the surgical field: through a pre-made guide, real-time navigation, or a personally designed fragment.
The choice is determined by bone deficiency, the number of implants, and the proximity of anatomical structures.
Comparison of planning methods:
| Option | When suitable | Limitations |
|---|---|---|
| Surgical guide | Single defects, sufficient bone volume | Position fixed in advance: tilt and depth cannot be changed during surgery |
| Dynamic navigation | Multiple implants, narrow bone, complex topography | Requires rigid jaw fixation and system calibration |
| Personal jaw model | Bone deficiency, augmentation, reconstructive cases | Longer preparation, more coordination with the surgeon |
When a guide is sufficient
If the bone allows placing the implant in the projected position without augmentation, the guide is the shortest path from plan to surgery.
The deviation remains within the tolerance set during planning, and the surgeon works according to the usual protocol and does not waste surgical time on equipment setup.
When a personal model is needed
With pronounced bone deficiency, one placement is not enough. We build a jaw model, check the position of the graft and future implants on it, and the same file goes into guide manufacturing — the plan and surgery do not diverge. The surgeon sees in advance the volume that needs to be restored.
In practice, approaches are combined: navigation for positioning a series, a guide for placement, a model for augmentation. We deliver the result in standard formats (tomography and surface model) so that the plan opens in the clinic's system, not only ours.
How we implement planning: from images to surgical protocol
Planning begins not with a model, but with data. If the image series is incomplete or there are metal artifacts in the area of interest, any further geometry is built on noise — and the error goes into the guide.
Audit of images and data
- Image audit. We cross-check tomography series and intraoral scans, verify completeness of the area, artifacts, and signs of patient movement. We request missing data before start, not during work.
- Bone tissue segmentation. We mark bone, mandibular nerve canal, sinuses, and roots of adjacent teeth as separate layers — boundaries remain visible and verifiable, not merged into a single volume.
- Model building. We transfer segments into a three-dimensional scene, position the implant along axes, and assess bone thickness around and occlusion.
Verification on model and handover
- Verification on test cases. We run the assembly on control sets with known results and compare discrepancy by reference points before giving the calculation to the surgeon.
- Handover of the plan to the surgeon. We provide the scene and printout, discuss access and controversial areas; if the axes do not match the clinical picture, we correct and reassemble.
- Surgical protocol. We fix positions, sizes, drilling sequence, and reference points for the guide — the operating room works according to this document.
After approval, the protocol and model go into guide manufacturing, and the original scene remains with the clinic as a reference: it is later used to verify the actual implant position.
What we hand over: models, schemes, regulations, and accesses
We hand over not a "working service" but a kit according to which the clinic manages the project itself or with our support.
Everything that describes the model logic, data, and accesses goes into your repository, and the handover is closed by acceptance: your specialist deploys the environment according to instructions and reproduces the result on a control set of cases.
Documentation and regulations
- Planning documentation — how the input set is assembled from images and impressions, which bone tissue parameters are considered, how the result is read, and where the model's applicability boundaries are.
- Implant placement schemes — position and axes in a format that opens in your planning system, together with templates for the surgical stage.
- Model update regulations — what triggers retraining, on which set the new version is tested, by which metrics it is accepted, and how to roll back to the previous one.
- Patient data access rights — roles, anonymization at input, log of image access, storage periods, and access revocation procedure.
- Environment configurations — deployment parameters, dependencies, hardware requirements, and installation procedure on the clinic side.
- Quality control protocol — control set of cases, methodology for comparison with actual surgical outcome, and acceptable discrepancy.
- Support procedure — support window, contact channel, incident owner, and response times.
Such a kit removes dependence on the contractor: the model, data, and rules for working with them remain with the clinic, not in someone else's closed access.
Example: how a class of failures was eliminated in planning based on a 3D model
The clinic came with a problem: the implant position was clarified directly in the operating room, and the discrepancy was detected only after placing the healing abutment.
Planning was done manually on CT slices, so the decision relied on the experience of a specific surgeon and was poorly replicated by other shifts.
What it was before
The plan was prepared in 2–4 days: a technician manually marked slices, built a 3D model, and printed a surgical guide. In 18% of cases, the position was corrected during surgery, and 9% of interventions ended in redoing — reprinting the guide and a second visit.
What it became after
We assembled a segmentation model: it identifies the root canal, mandibular nerve, and bone tissue boundary on CT slices and outputs a ready 3D model. The surgeon checks and confirms the plan — manual marking is gone, and the decision became reproducible in any shift.
| Indicator | Before | After |
|---|---|---|
| Plan preparation time | 2–4 days | 1 working day |
| Position correction in operating room | 18% of cases | 4% |
| Redos and repeat visits | 9% | 2% |
| Surgical guide reprinting | 12% | 1% |
Metrics are from the clinic's Реестра клинических исходов: comparison across 240 plans for the period before and after implementation.
The class of failures associated with positioning error has left regular practice — only isolated cases remain where the choice of point is limited by the anatomy itself.
How does a virtual plan differ from a physical guide?
The virtual surgical plan is a digital model: a segmented computed tomography image and the position of implants calculated taking into account bone density, adjacent roots, mandibular canal, and the future crown.
It answers the question "where and at what angle to place," but in the surgical field it exists only on the screen.
The physical surgical guide is an overlay printed or milled according to this plan. It seats on the teeth or mucosa and carries guide sleeves: the drill enters at a given angle and to a given depth. The guide turns coordinates into rigid mechanics and maintains the accuracy of plan transfer.
The difference is in the nature of the methods: the plan describes geometry and sequence, the guide physically limits the freedom of instrument movement.
The plan can be reassembled in minutes, the guide after manufacturing is unchangeable — a correction means re-production and postponement of surgery.
Hence the limitations. The guide needs stable support and sufficient mouth opening: with mobile mucosa, seating shifts and accuracy drops.
Closely standing teeth and narrow gaps interfere, and metal in the oral cavity creates artifacts on the image — the model error adds up with printing and sterilization errors.
Therefore, the plan is verified by fact: the postoperative image is superimposed on the original and the deviation of position and angle is examined. The quality criterion is not the fact of a manufactured guide, but falling within tolerance.
Where there is a single implant and calm anatomy, navigation according to the plan is sufficient; near a nerve or sinus, one cannot do without a guide.
Technical doubts: where AI errs and how we verify it
AI does not "see" anatomy — it reproduces the statistics of the markup on which it was trained. Segmentation accuracy depends on image uniformity, quality of reference standards, and acceptance metrics. Below are the doubts most often voiced by clinics and radiologists, and our answers.
Model quality control
We evaluate the model not by a single average figure: we calculate the Dice coefficient on a hold-out set separately for each anatomical zone and for each device, plus the share of cases sent for manual verification.
If boundaries diverge above the threshold, the system marks the image as uncertain and does not output it as a ready plan.
Different devices and image quality — will the model "drift"?
Yes, if trained on a single source. Therefore, the sample includes images from different devices, with metal artifacts and thin slices; the current nnU-Net branch covers part of this variance.
Before analysis, the image undergoes input control: if quality is low, the case goes to a doctor, not to the model.
Who is responsible for marking reference standards?
Reference standards are marked by radiologists according to a protocol, not by a contractor "by eye." Each case undergoes double-blind marking, disputed boundaries are reviewed by a third specialist, and discrepancies are logged.
The clinical side is responsible for the reference standard, and we are responsible for process reproducibility and dataset versioning.
What if the model errs on a specific patient?
The final decision remains with the doctor: the system provides boundaries and positioning options as a recommendation. The error can be localized — we save the input image, prediction, and model version, so the analysis proceeds step by step, not "blindly."
Let's discuss the task: what data is needed to start
The accuracy of implantation prognosis rests not on the algorithm, but on the quality of input data.
Therefore, the project begins not with model training, but with a short conversation: which clinical scenario we are modeling, which outcomes we consider target, and what has already been accumulated in the clinic.
Below is the minimum set that covers the assessment and allows assembling a work plan. If something is missing, say so directly: we will suggest what to replace it with or how to prepare it without loss of quality.
- CBCT or panoramic images in the original format (DICOM or a series of slices), without compression and screenshots.
- Intraoral scans or scanned jaw models — STL, PLY, or an archive from the office.
- Photo protocol and bite data, if the scenario includes an aesthetic part and crown positioning.
- Anonymized patient histories: medical history, comorbidities, risk factors, and complications.
- Outcome marking: what was considered success after the control period — engraftment, bone tissue condition, rejection.
- Volume of accumulated data: how many cases there are, in what form they are stored, and how quickly they are exported.
- Data restrictions: where records are currently stored, whether there is a ban on taking them outside the perimeter, and who grants access.
Describe the task — we will return with an assessment and work plan: what is needed first, what can be prepared later, and at what volume it makes sense to launch the first version of the model.