20% off on annual plans this month
code SMILE2020% off on annual plans this month
code SMILE20Obturation length, fill density, and apical seal — the three variables AI checks on every post-treatment radiograph, and what happens on the film when one of them doesn't add up.
What this article covers
A breakdown of how AI grades a completed root canal on a post-obturation X-ray — the three radiographic markers it measures, where the published accuracy numbers come from, and why apical seal is the one variable no AI can see directly.
A root canal looks done the moment the access cavity gets sealed. Whether it actually holds up gets decided earlier than that — at the point where the gutta-percha stops.
Underfilled canals fail more often than adequately filled ones, and overfilled canals sit somewhere in between — this is well-established in endodontic outcomes literature. Neither result is where a completed treatment should land. Both trace back to the same three-variable check a radiologist, or now an algorithm, runs on the post-op film.
That check has a name in the literature — technical quality assessment — and it's been done manually for decades, usually by an endodontist eyeballing a periapical X-ray and scoring length, density, and taper on a rough scale. The problem was never that dentists don't know what to look for. It's that two experienced clinicians grading the same film routinely land on different scores. A 2026 systematic review pooling the evidence on AI for this exact task found sensitivity for obturation quality assessment ranging from 0.79 to 1.00 and specificity from 0.99 to 1.00 — a wide range that itself says something about how unstandardized this field still is, even as it matures fast.
The three-variable check
Every AI system built for this task evaluates the same three things a manual radiographic review would: obturation length (how close the fill sits to the radiographic apex), fill density (whether the material is homogeneous or shows voids), and apical seal — inferred from how cleanly the first two line up near the terminus of the canal.
Not all three are equally easy to grade from a flat X-ray. Length is the most straightforward. Density takes more careful image processing. Seal, as it turns out, can't really be measured at all — only inferred.
Most classification systems, AI or human, use the same cutoff: a fill terminating within 0 to 2mm of the radiographic apex counts as adequate. Short of that line, it's underfilled. Past it — material visibly beyond the apex — it's overfilled. This is the parameter AI handles most reliably. A 2025 object-detection framework built specifically to classify RCF length on periapical radiographs reported an mAP50 of 87.9%, and the 2026 systematic review covering the field found this to be the most consistently well-performing sub-task across the studies it pooled.
That's the length variable handled reasonably well. Density and seal are harder.
A void inside the fill doesn't announce itself. On a periapical X-ray it shows up as a faint radiolucent gap inside otherwise dense gutta-percha — easy to miss on a monitor, easier to miss on a printed film. Models trained for this task flag density inconsistency by scanning pixel intensity along the length of the canal, looking for drops that shouldn't be there if the fill were homogeneous.
This matters more, clinically, than a length error does. A canal filled to the perfect distance from the apex with a void halfway up still leaves a channel for bacteria to recolonize.
Here's the part worth being upfront about: no 2D periapical radiograph shows apical seal directly. Seal is a three-dimensional property — whether the filling material has adapted tightly against the canal walls right at the terminus, with no microscopic gap for fluid or bacteria to travel through. A flat X-ray can't resolve that.
What AI, and human graders, actually do is infer seal quality from length and density together. If the fill terminates in the right zone and shows uniform density right up to that point, seal is assumed adequate. It's a proxy, not a direct measurement — worth remembering the next time an AI-generated quality score gets read at face value.
0.79–1.00
Sensitivity Range
AI Obturation Quality Assessment
Systematic review, 2026
0.99–1.00
Specificity Range
AI Obturation Quality Assessment
Systematic review, 2026
87.9%
mAP50
RCF Length Classification (YOLOv11)
RCFLA-YOLO, BMC Medical Education, 2025
How the four sub-scores are defined
Methodology noteLength
Distance from the fill terminus to the radiographic apex. 0–2mm = adequate. Short of the apex = underfilled. Beyond it = overfilled.
Density
Pixel-intensity uniformity along the canal. Radiolucent breaks in an otherwise dense fill get flagged as voids.
Taper
Whether the fill narrows smoothly toward the apex, mirroring the shape of the canal preparation.
Seal
An inferred composite of length and density. Not independently visible on a 2D film — this is the one score that's a proxy, not a direct read.
Manual radiographic scoring has a known weakness: two trained clinicians grading the same film can land on different scores often enough that inter-observer agreement gets reported as a standard metric in the literature. That's the gap AI is closing — not because a model understands endodontics better than a specialist does, but because it applies the same threshold every single time. A 2025 review pulling together ten separate studies on AI in endodontic diagnosis and treatment evaluation — covering CNNs, U-Net segmentation, YOLOv5, and even large language models like GPT-4 applied to radiograph interpretation — found diagnostic accuracy ranging from 75% to 99%, with sensitivity and specificity frequently above 80%.
One 2025 study built a YOLOv5 network on just over 1,000 periapical radiographs, then tested it against an external dataset of 500 more films from a different center — the harder test, since the model had never seen that clinic's machines or patient population. On that external set, it outperformed an inexperienced endodontist on overall grading accuracy. Not a specialist. An inexperienced one. Which is a fair indicator of where the real value shows up: catching the marginal case a newer clinician might grade inconsistently, not replacing an experienced eye.
| Parameter | What AI measures | Reported performance | Consequence if missed |
|---|---|---|---|
| Obturation length | Distance from fill terminus to apex | 87.9% mAP501 | Most reliably graded sub-task per 2026 review2 |
| Fill density | Void / homogeneity detection along canal | 0.79–1.00 sensitivity2 | Voids create microleakage channels |
| Overfill vs. underfill | Extrusion beyond apex vs. short fill | 77.5% precision3 | Both linked to lower long-term success |
| Apical seal | Inferred composite (length + density) | Not independently measurable | 3D leakage invisible on a 2D film |
Superscript numbers correspond to the reference list below. Figures represent findings from individual studies and a pooled systematic review, not a single unified benchmark; methods and datasets vary between the underlying papers.
AI flags, the dentist decides
Every system built for this task operates as decision support — it flags a case as needing review, it doesn't issue a final grade that goes in the chart. A flagged underfilled canal might still be clinically fine if the apex was calcified and the working length was adjusted on purpose; only the treating dentist has that context. Medecro's approach follows the same principle: flag, don't diagnose.
Everything above sounds tidier than it is in an actual clinic. A few things worth knowing before treating an AI obturation score as gospel:
2D projection distortion. Angulation errors — cone-cutting, elongation, foreshortening — change how long a fill appears on the film even when the actual length hasn't changed. A model trained on well-angulated images can misjudge length on a poorly shot radiograph, the same way a human grader would.
Seal is inferred, not measured. As covered above, no periapical X-ray shows true apical adaptation. A flagged "adequate seal" score is really a length-plus-density proxy. CBCT gets closer to a real 3D seal assessment, but most clinics don't shoot CBCT for routine post-treatment follow-up.
Radiographic success isn't clinical success. A technically well-obturated canal — right length, dense fill, clean taper — can still fail if a canal was missed entirely, if coronal leakage develops from a delayed restoration, or if the original infection wasn't adequately debrided before filling. Grading the fill says nothing about what happened during cleaning and shaping.
Validation on Indian datasets is still limited. Most of the published models above were trained and tested on patient populations and radiograph equipment from East Asian and European centers. Anatomical variation and equipment differences mean a model validated abroad needs local validation before its accuracy numbers can be trusted at face value in an Indian clinic — one reason Medecro's AI X-Ray Analyzer includes validation protocols built for Indian clinical settings.
The field is still actively validating itself. A clinical trial registered in 2025 is currently comparing AI-assessed obturation quality — length, density, taper, and coronal seal — against CBCT and conventional periapical radiography, with endodontist assessment as the reference standard. That a study this specific is still recruiting tells you the evidence base, while growing fast, isn't fully settled yet.
Regulatory status varies. India's CDSCO framework for AI-based software as a medical device is still developing. An obturation-quality AI feature running as decision support sits in a different regulatory category from one issuing an autonomous pass or fail — worth knowing before treating an AI grade as a formal clinical record.
The practical version: treat an AI obturation flag the way you'd treat a second opinion from a colleague standing over your shoulder — worth a second look, not a verdict. Clinics getting the most out of this run it as a standard post-op check on every RCT film, not just the ones that already look questionable. That's where it actually catches something.
What changes once it's routine is more interesting than the flag itself. A flagged film doesn't mean redo the treatment. It means look again — and often that second look confirms the fill is fine, the flag was an angulation artifact or an intentionally short working length. What it does reliably do is stop the marginal case from getting signed off without a second glance. Given how wide the sensitivity range still is across published models — 0.79 to 1.00 — that second glance is doing real work.
Medecro AI X-Ray Analyzer
Medecro's AI X-Ray Analyzer flags length, density, and taper on every root canal treatment radiograph — with confidence scoring and one-click override, inside the workflow you already use. No separate login, no standalone app.
Book a Demo — See It LiveObturation length, fill density, and apical seal — the AI evaluates all three on the post-treatment radiograph, flagging under-filled or over-extended cases for the treating dentist to review. Length is measured directly against the radiographic apex, density is checked for voids along the canal, and apical seal is inferred from how cleanly the first two line up near the terminus, since seal itself isn't directly visible on a 2D film.
Sources & references
