The May 2019 edition of The Atlantic magazine contained an article titled “The truth about dentistry”. In it, the author visualised dentists—not a particular dentist but dentists in the abstract—as sinister authority figures looming over the helpless patient’s recumbent form, drill in hand. Mistrust permeated the scene like swamp fog. “When he points at spectral smudges on an X-ray,” the author pleads, “how are we to know what’s true?”

Then there was the Dustin Hoffman movie Marathon Man with its Nazi dentist-cum-torturer, and the famous—or, if you’re a dentist, notorious—1997 Reader’s Digest article by a writer who visited 50 dentists in 28 states, picking them at random out of the Yellow Pages, and was given treatment plans ranging in cost from under US$500 to nearly US$30,000. That one really hit a nerve, so to speak.

Dentists have had their share of bad rap, but still, the experience of the Reader’s Digest writer was probably not terribly far from the truth. It was borne out, with eerie accuracy, by a 2021 Dental AI Council study intended to quantify the suspected inconsistencies in dental diagnosis and treatment. The same set of full-mouth radiographs was presented to 136 dentists, and they were asked to provide tooth-by-tooth diagnoses and a treatment plan. A person with confidence in the scientific basis of dentistry might naturally expect a limited amount of diversity among the responses and would assume that the commonalities would far outweigh the differences. Not so. Not once did more than half of the participants agree about the diagnosis for a given tooth. The variety of estimated costs was almost comical, ranging from US$300 to US$36,000—figures strikingly similar to those cited by the Reader’s Digest author. Worse, the range of cost estimates did not present as a bell curve, the majority of responses clustered together and only a few outliers at the extremes. Instead, the distribution was more or less flat; the frequency of a cost estimate of US$1,000 was about the same as that of a cost estimate of US$10,000.

Other studies have found that dentists’ interpretation of radiographs—the very foundation of diagnosis—was far from reliable. Estimates of cavity depth and recognition of radiolucencies were wrong as often as they were right. In another study, three dentists examined several thousand radiographs; their interpretations were in full agreement only 4% of the time.

Houston, we have a problem

How should we account for this lack of precision in a medical field? Is it due to dishonesty? To greed? To variations in skill? To honest differences of opinion? Whatever the reason, it gives dentistry a bad name. But there is a remedy. It comes in the form of a powerful new technology that is already transforming many aspects of our lives: artificial intelligence, or AI for short.

AI is an umbrella term covering a wide range of computing techniques. They range from “general AI”—intelligence indistinguishable from that of a human being, in all circumstances—to “narrow AI”, specialised programs whose expertise is limited to a particular class of problem. Most make use of a programming technique called a “neural network” by loose analogy to the structure of the human brain, and all have in common the property of trainability. They learn by taking in vast amounts of data of a certain type—say, photographs of faces or samples of text—and extracting commonalities. Once trained, an AI program can pick out a particular face in a crowd or write an essay or a love poem as well as or better than you can.

General AI is the darling of science fiction writers, but is very far from realisation. No AI system has anything like the broad knowledge of all aspects of the world that a human being has, and so, for the time being at least, we do not have to worry about being taken over by independent-minded and malevolent robots like the notorious HAL of 2001: A Space Odyssey. Even the comparatively limited task of safely operating a car in an urban environment has not yet been mastered, despite years of effort and oceans of investment.

“Once trained, an AI program can pick out a particular face in a crowd or write an essay or a love poem as well as or better than you can.”

Narrow AIs, however, already easily match or surpass human abilities, and they have become the tools of choice for performing many exacting tasks. Many of these involve computer vision, the analysis and recognition of objects or imagery. More than a decade ago, it was found that a trained AI could recognise and categorise nodules in radiographs of cancer patients’ lungs as accurately as a panel of oncologists could, and much faster. Computer vision and AI are now familiar parts of the oncological toolkit, and they are being applied to a widening array of medical fields. One of those is dentistry.

Dentists are in an excellent position to take full advantage of AI. There exists, to start with, a virtually limitless supply of dental radiographs for training. The radiographic image is the coin of the realm in dentistry; patients are accustomed to having their pathologies explained to them with reference to the “spectral smudges on an X-ray” evoked by The Atlantic’s reporter. The range of pathologies to be detected is relatively narrow, and the AI program can not only identify them but also quantify them with greater than human precision. The dental radiograph is, therefore, an ideal application for the sharp focus of narrow AI.

The second opinion—so to speak—provided by an AI program is directly valuable to the practitioner. The computer is hypersensitive to subtle greyscale gradations; it may detect something the human reader has overlooked. More importantly, it is never tired, distracted or rushed and so is not prone to the types of mistakes and oversights that people routinely make simply because they are human. The AI program may in many cases simply duplicate the perceptions of the human, in which case nothing is gained but confirmation, but it may add information overlooked by the human or differ in its interpretation, leading to a re-examination and re-evaluation of the evidence.

Even if these benefits may seem minor to an experienced practitioner confident in his or her abilities, there is another side of the AI experience to consider: the patient’s. The results of the AI program’s analysis are presented to the patient in vivid, intuitively understandable form. The radiograph no longer consists merely of spectral smudges, but has become graphically compelling, having highlighted areas, colour-coded outlines and explanatory labels. For a patient, the enhanced display conveys a heightened sense of precision, clarity and objectivity. The diagnosis is no longer just the opinion of one person, whom a cynic might suspect of ulterior motives. It need not be taken on faith; it is supported by the unbiased authority of a digital computer.

“The computer is hypersensitive to subtle greyscale gradations; it may detect something the human reader has overlooked.”

While the graphic presentation of a computed analysis may impress a patient as something more than human, the practitioner should be aware that the AI program is an assistant, not a supervisor. Even though the accuracy of AI’s radiographic analyses in various medical fields has been shown to be indistinguishable from that of human interpreters, the AI program actually knows much less about teeth (or lungs or livers) than the trained and experienced practitioner does. What it does know, and knows very well, is how a large number of specialists have interpreted a very large number of radiographs. Its findings are, in effect, those that hundreds or thousands of dentists would make if they were to vote on the content of a given radiograph. Where there is not unanimous agreement, majority opinion prevails, or findings are presented in terms of probabilities. The practitioner using the AI program remains entirely free to form a different opinion or to disregard the advice the program gives, but has the benefit of knowing what a large group of peers would have made of the radiograph in question.

The most significant impact of dental AI, however, is not that it necessarily brings a superhuman level of certainty to the data upon which diagnoses are based—although in most cases it may—but that it provides, for the first time, an objective and universally accessible standard of reference. Objective standards are precisely the thing that dentistry has lacked in the past, and their absence has given rise to suspicions about the candour and consistency of dental diagnoses. Look at the Reader’s Digest writer: guided only by a phone book, he collected a bewilderingly large variety of diagnoses. If he had visited only dental offices using an AI assistant, he would have been given a much smaller variety, and the differences would have been due to small variations among the radiographs made by different practices rather than to the whims of individual dentists or the immediate financial needs besetting them.

Consistency is not the only thing AI brings to dentistry. It also provides support for insurance claims and facilitates record-keeping, tracking of patients’ dental health and comparison of performance among multiple practices in an organisation. It trains dentists at the same time as dentists train it. In the future, it may reveal connections between dental health and general health that we do not now suspect.

Those are some of the collateral benefits. Above all, however, AI will give patients the reassurance of knowing that the condition of their teeth is not merely a matter of opinion.