What CBCT Deep Learning Can — and Cannot — Do

Cone-beam computed tomography (CBCT) has become part of routine imaging in endodontics, implantology, and complex esthetic planning. Over the last five years, a growing body of literature has proposed deep-learning models to segment anatomy, detect periapical lesions, classify root morphology, and support decision-making around these images. The volume of publication has grown considerably faster than the volume of clinically validated tools.

The question worth asking is not whether artificial intelligence “works” on CBCT, but where the current evidence is genuinely informative for a practicing dentist, and where it is not. This article summarises how I read the current literature after completing a PRISMA-aligned narrative synthesis of United States and European studies from 2021 through 2026.

What the evidence reasonably supports

Several tasks show consistent, reproducible model performance across independent datasets, even when the datasets are relatively small.

• Segmentation of well-defined anatomy. Mandibular canal tracing, maxillary sinus outlines, and tooth-level segmentation are the most mature tasks. Reported performance is stable across scanners in multiple studies and clinically useful as a starting point that a clinician still reviews.
• Detection of periapical radiolucencies as a secondary reader. Multiple studies report sensitivity comparable to experienced clinicians on curated test sets. These models are best interpreted as prompts for closer human review, not as diagnoses.
• Landmark localisation for planning. Automated landmarking supports implant and orthodontic planning steps, and shortens preparation time when the clinician retains final placement authority.

Where the evidence is fragile

Reported metrics can look impressive in a single study and then drop substantially when the same model is used on scans from a different device, protocol, or population.

• Small, single-centre training sets remain the norm. Many models are trained on data from one institution and one CBCT unit. External validation is often absent.
• Class imbalance — for example, few positive periapical cases — inflates headline accuracy while masking real error patterns. Sensitivity, specificity, and calibration matter far more than a single “accuracy” figure.
• Scanner variability is not cosmetic. Field of view, voxel size, reconstruction algorithm, and artefacts change what the network “sees.”
• Ground truth in CBCT studies is frequently defined by consensus of a small number of readers, not by histology or long-term follow-up. This limits how far the model has been shown to generalise.

What remains under clinician responsibility

Regardless of model performance, several elements are not appropriate to delegate to a deep-learning tool in current practice.

1. Diagnosis. A model output is a data point, not a diagnosis. It is read together with history, clinical examination, other imaging, and clinical context.
2. Treatment planning. Sequence, materials, prognosis, and alternatives remain a clinician decision.
3. Informed consent. Patients are informed by a clinician who can answer their questions, not by a probability score.
4. Regulatory and privacy compliance. Whether a specific tool is cleared, how patient data is stored, and how it is transmitted are governance questions, not research questions.

A practical reading posture

When I read a new CBCT-and-AI paper, I focus on four questions before I focus on the headline metric:

• How many centres and how many scanners contributed data?
• Was there any external test set, and how was it curated?
• How was ground truth defined, and by whom?
• Are sensitivity, specificity, and calibration reported — not only accuracy?

These questions do not require access to code or original data. They tend to separate studies that are useful for clinical thinking from studies that read impressively in a slide deck but say very little about a real clinic.

Closing note

Deep learning is already a useful adjunct for CBCT interpretation in narrow, well-defined tasks. It is not currently a substitute for the trained reading of a dentist who understands what the image, the patient, and the plan actually require. A responsible use of these tools is precisely what makes them safe to keep integrating.