Beyond the Apex: Navigating AI in Endodontic Diagnosis, Education and Ethics

By Ashley L. Madern, DMD, MS

A student once asked after lecture, “If AI can see more pixels and grayscale than my eye can, shouldn’t I trust it when it identifies a periapical radiolucency?”

This question captures the complexity of AI in both clinical and educational contexts.

AI systems are being trained to detect patterns humans may overlook, and while they can identify radiographic findings well, identification is not the same as interpretation. AI can recognize patterns, but it cannot contextualize them with clinical reasoning, judgment, and experience.

The question also highlights another concern: newer clinicians are more likely to trust the computer over themselves. Confidence comes from experience, repetition, and occasionally being wrong. If learners allow AI to replace, rather than augment, development of diagnostic skills, they risk becoming dependent on it before building interpretive independence.

AI is no longer a future consideration in endodontics. It is here, and students are navigating it without formal guidance.

AI Has Already Entered Endodontics

Artificial intelligence is quickly expanding across endodontics. Software exists for treatment planning, guided access, and surgical guide creation. Deep learning models can detect periapical pathosis on CBCT scans, segment pulp space, classify C-shaped canals, and detect vertical root fractures on intraoral and cone beam imaging. These are not theoretical concepts; they are peer-reviewed applications with growing clinical relevance. For educators, however, the most important thing to teach is not how these tools work, but how to evaluate them.

Numbers Are Not the Whole Story

When a vendor advertises high sensitivity or accuracy, the first question should be: compared to what?

AI performance is measured against a human-labeled reference standard known as “ground truth.” If clinicians labeling the data disagree with one another or practice with different diagnostic philosophies, those biases become embedded into the algorithm itself.

A study reporting 92% sensitivity for periapical lesion detection only matters if we understand who labeled the data, how consistently they agreed, and whether the system was validated on imaging systems and patient populations similar to our own. If an AI system consistently disagrees with your interpretation, that disagreement may reflect differences in training data, imaging parameters, or diagnostic philosophy. In that situation, AI may undermine rather than support clinical judgement.

The regulatory landscape adds another layer of nuance. Most dental imaging AI products enter the market through the FDA’s 510(k) pathway, which requires demonstrating substantial equivalence to a prior device rather than proving improved patient outcomes. Manufacturers largely submit their own performance data, and independent external validation is not required. Additionally, algorithms evolve after clearance and without mandatory oversight of post-market changes.

That does not mean these tools lack value, it means they should be critically evaluated rather than accepted as infallible black boxes where we input an image and receive an answer without knowing how it was trained, the ground truth, whether it altered data (which AI can do), etc. Because any lack of transparency can lead to over-trusting the interpretation.

Automation Bias Is the Real Clinical Risk

One of the greatest dangers with AI is not incorrect output but instead overreliance and automation bias. Automation bias is the tendency to favor suggestions from an automated system even when contradictory evidence exists. In healthcare, this can quietly influence clinical decision-making.

Imagine driving in an unfamiliar city toward the airport when you see a sign that says, “Airport 1 mile.” However, your GPS says, “no, keep going.” Soon, another sign says, “Airport next exit.” But your GPS says “no, don’t take that exit with the giant airplane symbol, keep driving.” Would you listen to the signs or the GPS? If you say you would trust the AI then you fell victim to automation bias and probably missed your flight.

The same phenomenon occurs in radiology. A radiographic overlay may confidently flag a region as pathology, and a newer clinician may hesitate to challenge it even when the evidence does not fit.

Our Curriculum Has Not Caught Up

Most dental schools still provide little formal education on AI evaluation, implementation, or ethics. At the same time, many faculty members were trained before AI tools entered clinical workflows and may feel uncomfortable teaching concepts they themselves are still learning. Today’s students are entering a professional environment fundamentally different from the one many educators are trained in. Additionally, students are accustomed to immediate answers from technology and are more likely to trust those answers without questioning them. We have all experienced this ourselves: asking AI or the internet a question and accepting the response at face value.

General-purpose AI systems can produce inaccurate information, unsupported conclusions, or fabricated citations. More specialized medical AI systems may perform significantly better, but clinicians must be able to critically evaluate what they are using, how it was trained, and whether it is appropriate for clinical decision-making. As educators, we cannot simply teach students to use AI; we must teach them how to question it.

Data Bias: Who Trained it and on Whom?

AI systems are only as good as the data used to train them. If a model is trained predominantly on one demographic population, imaging system, or diagnostic philosophy, performance will decline outside those conditions.

For example, will an AI system correctly distinguish periapical cemento-osseous dysplasia (PCOD) from inflammatory periapical disease in populations where PCOD is more prevalent? Was the model trained on ideal CBCT scans from academic centers, and will it perform equally well on scans acquired with poorly calibrated equipment or inconsistent imaging protocols? Equipment bias, population bias, and training bias all matter.

Even image acquisition settings influence outcomes. Differences in kVp, resolution, calibration, artifacts, or reconstruction protocols may affect how an algorithm performs in private practice compared with the environment in which it was developed.

Clinicians should evaluate AI systems using their own previously diagnosed cases and ask an important question: does this system consistently agree with sound clinical judgment in my environment? If not, that discrepancy deserves attention.

The Clinician Remains Responsible

The governing principle should remain simple: the dentist decides; AI suggests.

AI can identify areas of concern, assist with workflows, and improve efficiency. But it does not hold a dental license, cannot appear before a dental board, and cannot assume medicolegal responsibility for irreversible treatment decisions. When AI contributes to a clinical error, liability remains with the clinician. That reality makes documentation increasingly important. If AI is used to help diagnose, clinicians should document that the findings were independently evaluated and clinically correlated. Most importantly, clinicians must resist the temptation to let AI become the primary driver of suspicion. Clinical judgment should originate from examination, interpretation, and reasoning, not from an algorithmic overlay.

What We Owe Our Learners

Students entering practice today will evaluate, purchase, and use AI tools often with little formal training in how to do so. Endodontic education is uniquely positioned to address that gap.

The important questions are practical ones:

• Was it externally validated?
• What is the false positive rate?
• Who established the ground truth?
• Was it trained on imaging systems and patient populations similar to mine?

AI is a reason to deepen clinical expertise, strengthen critical thinking, and reinforce the importance of independent diagnostic judgment. Ultimately, the future of AI in dentistry is not to replace clinical judgement, it is to make good clinicians even better.

Disclosure: The author has no relevant financial relationships with any commercial AI product or company mentioned in this article.

About the Author: Ashley L. Madern, DMD, MS, is a Clinical Assistant Professor and Oral & Maxillofacial Radiologist at Midwestern University College of Dental Medicine, where she teaches oral radiology and oral health sciences.