CBCT segmentation quality: what to review beyond the render
A polished render is not a segmentation-quality measurement. Understanding a dental model requires identifying its source, checking its scope, and distinguishing presentation changes from the underlying delineation. This guide proposes a technical and editorial review of demonstration-only representations, not a clinical validation protocol.
The goal is to ask better questions when examining technology: what is represented, where the limits are, and what the image cannot establish.
Start with scope, not appearance
Before judging a surface, ask what it was intended to represent. An object can look clean while failing to match the scope you expected. A demonstration of an arch's general form, for example, is not the same request as separately represented teeth.
Write the objective in one sentence: “Show these structures for this demonstration purpose.” Review the scene against that objective rather than assuming that every model supports every possible use.
Identify exclusions as well. A missing object may simply be outside the request. Its absence from the model must not automatically be interpreted as absence from the scan.
A smooth surface can leave important questions unanswered
Segmentation tools support region editing and smoothing operations. The final surface should therefore be understood as a processing result, not a neutral copy of the source volume. S08
As an editorial practice, keep a note of relevant operations. “Smoothed for visualization” describes a presentation decision. “Corrected the anatomy” makes a much stronger claim and should not be used without evidence supporting it.
The same principle applies to closing holes, simplifying geometry, or removing small components. A demonstration should identify the version being shown and acknowledge relevant transformations between source and output.
Where to focus visual review
We recommend examining object separations, surface ends, and regions the operator has identified as difficult to delineate. These review points do not certify a model; they make the discussion of its limitations more explicit.
When source information is affected by artifacts or unclear boundaries, the most flattering angle is not enough. Dental segmentation research specifically addresses the difficulty of separating teeth from neighboring structures. A polished example does not eliminate that underlying challenge. S14
For a useful demonstration, a comparison of recognizable views with a brief explanation is preferable to an animation that leaves the viewer unsure what is being shown.
How to read an algorithm's published metrics
Research may report overlap measures such as Dice and measures of surface distance. A CBCT nnU-Net study, for example, reports performance within the ToothFairy2 evaluation. That is evidence about the evaluated method and data, not a certificate that transfers to any software or scan. S09
Ask which data were evaluated, which structures were included, and what reference was used. Then distinguish those findings from claims about ASTHECA. Without a documented evaluation of our own process, we should not publish accuracy percentages, error margins, or claims of superiority over other services.
Metrics belong in an explanation of methods and evidence, not merely as a persuasive number in a headline.
A minimal record for a demonstration
We recommend recording the example's purpose, source modality, included structures, processing version, and known limitations. When information is missing, say it is undocumented rather than inferring it from the file's appearance.
Distinguish technical review from clinical evaluation. Checking that a model opens, its labels are legible, and its caption matches the scene does not validate it for patient care.
This record can be short. Its value lies in answering questions that the render cannot answer on its own.
Frequently asked questions
Does more visible detail mean greater accuracy?
That should not be used as an acceptance rule. Detail and visual finish do not document each surface's origin or its correspondence with an appropriate reference.
Does manual review make a file clinically validated?
No. “Reviewed” needs to specify what was reviewed, by whom, and for which purpose. ASTHECA's current scope remains demonstration-only. S13
What should I do when a region is unclear?
Request a description of the limitation rather than presenting it as a finding. Clinical questions must be resolved through the appropriate professional evaluation, outside this demonstration.
Technology that can be explained
Read how to interpret teeth and roots in a 3D model and ASTHECA's existing guide to connecting slices and reconstructions.
Explore ASTHECA's dental representations. Scope and delivery are coordinated before processing for requests from the United States and Mexico. Outputs are not for diagnosis, treatment planning, or guiding procedures.
Sources and references
- S08. 3D Slicer — Segment editor — 3D Slicer.
- S09. Scaling nnU-Net for CBCT Segmentation — Isensee et al. / arXiv, 2024.
- S13. ASTHECA — Your scan, in 3D — ASTHECA, U.S. version consulted 2026-09-06.
- S14. A novel deep learning system for multi-class tooth segmentation and classification on cone beam computed tomography. A validation study — Shaheen et al. / research text published by Relu.