AI and dental x-rays: what the evidence shows so far

This post summarizes published research for general education by Copeland Consulting Services. It is not clinical advice. It names no products.

What does AI do with a dental radiograph?

It looks for patterns that match labeled examples, such as a cavity or a bone defect. A 2023 review in Frontiers in Dental Medicine explains the base idea: in a two-dimensional radiograph, each pixel has an intensity that represents density, and an algorithm learns patterns from those values to segment teeth, detect caries and make predictions.

The 2023 review also describes how such a model is built. Clinicians label images by hand. Computer scientists prepare the data and the algorithm. Then both groups check the results together. Labeling takes time, and it shapes what the model can find.

For the technology itself, see artificial intelligence in dentistry and digital radiography.

What have studies reported?

The reviews list results for caries, periapical lesions, and bone loss. The table gives examples as the reviews report them. Each figure comes from a single study within a review, under that study’s own conditions.

The 2023 review adds a cost finding. It says one study compared AI with dentists for proximal caries detection and found the AI “more effective and less costly.” That is one reported study, and its setting may differ from yours.

Do not read these numbers as a promise. They come from different datasets, designs and definitions, so they cannot be compared side by side.

Why is reading radiographs hard for AI?

Radiographs hide small and early changes. The 2025 review lists several problems:

• Noise from patient movement, equipment quality and operator skill can create artifacts.
• Carious lesions make up on average only 1.5% of a panoramic image, and shallow caries less than 0.5%. This small size is hard for neural networks.
• About 40% demineralization of hard tissue is needed before a lesion shows on a radiograph.

That last point cuts both ways. Early lesions are faint for a human reader and for software. The 2023 review notes that early lesions are hard to detect with deep fissures, tight interproximal contacts and secondary lesions, and that many lesions are found only at an advanced stage.

Do study results carry over to a real office?

Not automatically. The 2025 review says the field is “in its nascent stages,” with “high precision in experimental settings” but open questions about “clinical accuracy and reliability.”

It also says models that depend on specific data “often underperform in new environments,” because image quality and noise differ between devices. The 2023 review makes a related point. Data from one country or region can make a model “precise but not accurate” for places with a different health system.

Other limits named by the reviews:

1. Labeled data is scarce, because manual labeling is slow.
2. Privacy rules make it harder to build large, standard datasets.
3. Learning algorithms can carry bias.
4. The “black box” problem: some tools cannot explain a result.
5. Legal risk tied to malpractice and data privacy.

What does this mean for a practice?

It means AI reading fits as a second look, with the clinician in charge. The 2025 review says AI systems should “complement” clinical decision-making “without supplanting the nuanced judgment of experienced clinicians.”

Before adopting any tool for radiographs, a practice can ask:

1. On what images was the tool trained and tested, and do they resemble the practice’s images?
2. Was it tested on images from other offices and devices?
3. What are its false positive and false negative rates?
4. How does it show its reasoning?
5. Who is responsible for the final read, and how is that documented?
6. Does the vendor sign a business associate agreement?

What do the reviews say about other image types?

The reviews cover more than bitewing and periapical films. The 2025 review reports AI work on panoramic images for caries segmentation and bone loss, on near-infrared transillumination images, and on photographs from standard phone cameras. It also reports studies on optical coherence tomography.

These other image types matter for radiograph decisions. Some of the methods use no ionizing radiation, which is why researchers study them. The 2025 review notes that, despite the relatively low dose from X-rays, “there remains a risk to patients.”

The 2023 review lists tooth-level tasks across panoramic, periapical and cephalometric images. It covers tooth segmentation, caries detection, periodontal bone loss and apical lesion detection. In each area, the same caveat applies: the data and the algorithm decide what the tool can find, and the clinician still decides what to do with the finding.

Teams should also match any tool to their informed consent and record-keeping routines. Ask your attorney how.

For the wider evidence, see AI in dentistry: what the research says. For the radiation side of imaging, see cone beam CT and radiation safety.

Next step

The AI in the dental practice guide covers uses, evidence and risks. To discuss a related project, contact Dr. Copeland about consulting.