Seeing More Is Not the Same as Seeing Better: AI in Dental Radiology

Radiology is where AI in dentistry has the longest record, and where the gap between a promising result and a dependable clinical tool is easiest to see. Colleagues rarely ask me whether the software is accurate. They ask whether they can safely rely on it. My answer, the better tools are useful as a second reader, and that the first reading remains the dentist’s.
Caries detection
Caries detection is the most studied application, and the results vary widely. One review of proximal caries reported accuracy ranging from 28.5% to 100% across studies, so a statement like “AI detects caries” covers many different models, datasets and reference standards.
How dentists perform with the software matters more than how the software performs alone. In the ADEPT study, 23 dentists were randomised to read bitewings with or without an AI tool. Those using it found 75.8% of enamel-only lesions against 44.3% in the control group, and also flagged 14.6% of healthy surfaces against 3.7%. More lesions were found, and some of them were not disease. Reports on a separate randomised trial suggest that dentists treated enamel lesions more aggressively when AI was used. That point reaches me through a secondary summary, but it matches a risk I see. A lesion highlighted on a screen calls for risk assessment and a preventive plan before anyone picks up a bur.
Periodontal bone loss
Measuring radiographic bone loss by eye is slow and varies between observers, so automation is attractive. A systematic review of AI on panoramic and periapical films included 30 articles, of which ten were pooled. The pooled sensitivity was 87%, specificity 76% and accuracy 84%, with variation between studies. A specificity of 76% means that roughly a quarter of unaffected cases could be flagged, which matters if the output feeds a treatment plan.
I regard these tools as a measurement aid. Probing depths, bleeding, mobility, furcations and risk factors decide the diagnosis and stage, and none of them appear on the film.
Endodontic imaging
On two-dimensional images, AI can only work with what the image records. A narrative review notes that panoramic radiographs detect periapical lesions with a sensitivity of about 34%. A systematic review found that CBCT gives the highest diagnostic accuracy for periapical lesions, and that AI performed better on mandibular lesions than on maxillary ones.
For canal anatomy, the evidence is thinner. A systematic review of AI for accessory canals, mostly the second mesiobuccal canal on CBCT, found only seven studies. Sensitivity ranged from 80% to 100% and AUC from 0.57 to about 0.90, and the authors called the evidence limited. Software that prompts you to look harder for an MB2 can help. You still have to find it, with magnification, in the chamber.
CBCT and surgical planning
The strongest CBCT evidence concerns segmentation, in which the software outlines teeth, jaws and the mandibular canal. A systematic review concluded that deep learning models show high accuracy, comparable to expert manual segmentation. A review of canal localisation covered 1,288 scans compared with expert manual tracing and asked future studies to follow AI reporting guidelines, which suggests that methods across the field are uneven.
Comparisons between products are more sobering. A head-to-head study of five platforms reported that accuracy varied widely, with the largest errors in complex regions and near the mandibular and mental foramina. The version I found was hosted on the website of one of the companies involved, so its independence is uncertain, but the finding is plausible. The software is weakest where the anatomy is hardest, and those are the places where an error costs most. For implants and third molars, an automatically traced canal is a starting point to be checked slice by slice.
Cephalometric analysis
Orthodontics has some of the more reassuring figures. A meta-analysis from the Berlin group found that about 80% of landmarks were placed within 2 mm of the reference, while asking for more evidence on robustness and generalisability. A later review of 2D cephalograms reported a pooled error of 1.39 mm, under the 2 mm threshold, but a mean error of 1.0 to 5.8 mm on 3D CBCT images. Automated tracing takes under a minute, which is a real saving.
Not every landmark carries the same weight. A point 1.5 mm out may change nothing, while 3 mm on a key reference can shift an angle and with it the plan. I check the landmarks that drive the decision, and I am more cautious with automated 3D analysis than with 2D.
In conclusion
A good tool earns its place as a second reader. It can draw my eye to a surface I passed quickly, outline a structure faster than I would trace it, and give a younger colleague a consistent starting point.
It does not carry the diagnosis. A radiograph has to be justified before it is taken, taken well, and read against the history, risk and clinical findings. Software does not reduce the need to think about dose. A coloured overlay is also very persuasive when shown to a patient, which is a reason to use it sparingly. If I would not have raised a finding without the software, I should be able to explain why I now think it matters.
Should AI findings be recorded in the patient’s notes, even when the dentist disagrees with them?
Tell us in the comments.




Comments