September 3, 2026 From Detection to Diagnosis: AI-Assisted Radiograph Reporting
An artificial-intelligence system can look at a bitewing and, in a fraction of a second, place a neat luminous ring around a faint radiolucency between two molars. It is an impressive act, and it is a small one. The ring says here is something the model considers worth your attention. It does not say the something is caries rather than a cervical burnout shadow, does not weigh it against the patient’s history of remineralization, does not decide whether to watch it or drill it, and above all does not put a licensed name to that decision. Between the ring and the treatment plan sits a document – the report – and the whole discipline of using these tools honestly lives in understanding that the software writes only part of it, and never the part that matters most.
We have written elsewhere about how the machine draws the ring in the first place – the caries and bone-loss detection that turns a flat field of grey into flagged regions in How AI Reads a Bitewing, and the voxel-by-voxel naming of a three-dimensional volume in AI on CBCT. This piece is about what happens next: the translation of a finding into a diagnosis, and of a diagnosis into a signed record.
Two Different Verbs
To detect is to notice and localize: to say a region of an image differs from its surroundings in a way associated with disease. To diagnose is to interpret: to fold that finding into everything else known about the patient – symptoms, history, clinical exam, the other radiographs, the plausibility of the shadow being an artifact – and arrive at a named condition with consequences. The two are routinely spoken of as one act because a skilled clinician performs them in the same glance. They are not one act, and current dental AI is built almost entirely for the first of them.
The distinction is not pedantry. A detection model trained on thousands of annotated bitewings has learned the appearance of caries; it has learned nothing about this patient. It cannot know that the dark area it flagged sits under a restoration placed last year, that the patient reports no sensitivity, or that the identical shadow on the contralateral film was dismissed as burnout at the last visit. Detection is population knowledge applied to a pixel. Diagnosis is that pixel read in the context of a person. The report is where the second is layered onto the first.
The Anatomy of an AI-Assisted Report
A modern AI reporting pipeline does more than mark spots, and it is worth seeing the assembly. From the raw radiograph the system produces, for each candidate finding, a localization (where), a label drawn from its trained vocabulary (what class of thing), a confidence value (how sure), and increasingly a measurement – the depth of a lesion relative to the enamel-dentine junction, the percentage of bone loss from cementoenamel junction to crest, the proximity of a planned implant to the nerve canal. These are then composed into a structured document: findings grouped by tooth, ranked or filtered by confidence, rendered as an overlay on the image and as a written summary a patient or an insurer can read.
Platforms differ in how far down this chain they reach. Pearl’s Second Opinion and Overjet sit largely in detection and measurement, presenting quantified findings for the clinician to interpret. Diagnocat pushes furthest toward the finished document, generating structured radiology reports across dozens of conditions from 2D films and CBCT alike – the volumetric side of which we traced in Diagnocat on CBCT Volumes. But a generated report, however polished, is a draft. It becomes a diagnosis only when a clinician adopts it.
The Regulatory Line: CADe, Not CADx
Regulators drew this same distinction long before the current wave of tools, and the vocabulary is worth knowing because it governs what a vendor is actually allowed to claim. The FDA separates computer-assisted detection (CADe) – software that flags or marks regions of interest for a clinician to evaluate – from computer-assisted diagnosis (CADx) – software that characterizes a finding, offering an assessment of what it is or how likely it is to be disease. CADe is a pointer; CADx renders an opinion, and it is held to a far higher evidentiary bar.
Nearly every FDA-cleared dental imaging AI to date clears as a detection or measurement aid, a concurrent read that assists a dentist who remains the diagnostician of record. Pearl’s Second Opinion, for instance, is cleared to aid detection of numerous radiographic conditions – the human makes the call. The nuance matters internationally, too: Diagnocat’s broad multi-condition analyzer has been registered in the US database on an export-only basis, while its specific US clearance is for a CBCT visualization tool – not blanket authorization for autonomous diagnosis. Reading a vendor’s regulatory status precisely, rather than from its marketing, is the difference between deploying a tool within its cleared intent and quietly stepping outside it.
The Clinician in the Loop
Because the cleared intent is assistance, the human step is not decoration – it is the mechanism by which a probabilistic overlay becomes a medicolegal record. Mature systems present each finding as something to be accepted, rejected, or adjusted: the dentist confirms the flagged lesion, dismisses the burnout artifact, corrects the tooth number, edits the bone-loss measurement where the model clipped the crest. Only the adopted findings pass into the chart, and at that moment their authorship transfers. The report is no longer what the software proposed; it is what the clinician concluded, and the responsibility for it is entirely human.
This is also where the value of AI reporting is realized or squandered. Used well, the overlay is a tireless second reader that catches the interproximal lesion missed at the end of a long day and quantifies the bone loss more consistently than the eye. Used badly – waved through without genuine review – it launders the model’s errors into the patient’s permanent record under a licensed signature. The tool cannot tell the two uses apart. Only the clinician’s discipline can.
Where the Report Misleads
The failure modes of AI reporting are subtler than a missed lesion, because they arrive wearing the costume of thoroughness. The first is over-flagging: a model tuned for sensitivity marks every faint shadow, and a report dense with low-confidence findings buries the one that matters and trains the clinician to click past all of them. The second is automation bias – the well-documented human tendency to defer to a confident-looking machine, to accept a flag one would have dismissed unaided simply because the software drew a ring around it. The third is miscalibrated confidence: a stated ninety percent that does not correspond to being right nine times in ten, on this patient’s demographics, this sensor, this exposure.
Underneath all three sits the constraint we return to across this series: the report can only ever be as good as the image beneath it. A model asked to diagnose from an underexposed, cone-cut, or motion-blurred film will still produce a confident, tidily formatted report – and that composure is precisely the danger. A clean radiograph is not a nicety here; it is the precondition for the finding, the measurement, and the diagnosis all being about something real.
Future Developments
The trajectory is toward reports that are less a list of isolated flags and more a longitudinal argument: this lesion compared to its appearance eighteen months ago, this bone level tracked visit over visit, the AI narrating change rather than merely marking a single frame. As foundation models fold radiographic findings together with the periodontal chart, the clinical notes, and the intraoral scan, the generated draft will read more like a colleague’s considered impression than a pattern-matcher’s list. That is genuinely useful, and it sharpens rather than dissolves the central point. The more fluent the machine becomes at composing the report, the more the report will sound like a diagnosis – and the more it will matter that diagnosis remains an act a licensed clinician performs, in the context of a whole person, and signs. The software will get better at drafting. The authorship, and the judgement it carries, stays ours.
Sources & further reading:
No Comments