What Is AI-Assisted Dental X-ray Analysis?

Most systems are trained for a handful of well-defined tasks on a specific kind of image, such as panoramic or bitewing radiographs. Typical outputs include:

• Tooth detection and numbering, usually with the FDI two-digit system (for example 36 = lower left first molar).
• Existing dental work — fillings, crowns, root canal fillings, implants — which helps describe the current state of the mouth.
• Candidate findings — areas that may be consistent with, for example, decay, changes around a root tip or bone loss, flagged for professional attention.

What a system can recognise depends entirely on what it was trained and validated for. A tool trained on panoramic images may not be reliable on a photo of a periapical film, and vice versa.

Results are often shown as coloured boxes or outlines drawn over the X-ray. They answer the question “where?”, not “what for certain?”. An outline around a tooth surface means the software found a pattern there worth checking; the dentist then looks at the same area, compares it with other images and the clinical picture, and decides what it is.

Overlays are useful in conversation: they make it easier for a dentist to show a patient exactly which area they are discussing.

Many systems attach a number to each mark, often called a confidence or probability score. It reflects how strongly the image region resembles what the model learned. It is not:

• a measure of how severe or urgent a problem is;
• a guarantee that the finding is real;
• a diagnosis.

Machine-learning research has also shown that modern neural networks can be poorly calibrated — their scores may sound more certain than their real error rate justifies — unless they are specifically adjusted. That is one reason a high score still needs professional review, and a low score does not mean “nothing to see”.

A false positive is a mark on something that turns out to be normal — for example an overlapping shadow that looks like decay. A false negative is a real problem the software did not mark. Both are expected, for reasons such as:

• overlapping teeth or structures in a two-dimensional image;
• blur, low contrast, cropping or a phone photo taken at an angle;
• early or subtle changes that experts themselves may disagree on;
• images that differ from the training data (another type of X-ray, another machine, another population).

A systematic review of deep learning for caries detection found that study results vary widely and that many studies had methodological limitations; reporting checklists and new ADA validation standards aim to make such evidence more comparable. Before uploading, a few image-quality checks help reduce avoidable errors.

FDI recommends that dentists use AI as an assistance system and stay responsible for any diagnosis or treatment. In practice, a dentist reviews each candidate finding and either confirms it, rejects it or adds observations the software did not make. They may ask for a clinical examination or a different image before deciding — for instance a bitewing for suspected decay between teeth, or a periapical image to look at a root tip.

Only after that review does an observation become part of a diagnosis and, possibly, a treatment discussion. Our guide on reading an AI-assisted dental report explains how these stages appear on a report.