📄 medicine

Development and external validation of an AI pipeline for fully automated DMFT determination from panoramic radiographs

This study presents a deep-learning pipeline that achieves fully automated DMFT determination from panoramic radiographs with overall agreement comparable to that between human dentists, demonstrating strong performance for identifying missing and filled teeth while highlighting persistent challenges in detecting decayed teeth due to the limitations of panoramic imaging.

Original authors: Babak Saravi, Veronika Shavlokhova, Ashraf Türky, Daman Deep Singh, Felix Schrader, Andreas Vollmer, Norbert Kübler, Christoph Sproll, Lara Schorn

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Babak Saravi, Veronika Shavlokhova, Ashraf Türky, Daman Deep Singh, Felix Schrader, Andreas Vollmer, Norbert Kübler, Christoph Sproll, Lara Schorn

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your mouth is a busy city, and the DMFT index is the city's official report card. It counts three things:

Decayed teeth (the ones with cavities).
Missing teeth (the empty lots).
Filled teeth (the ones that have been repaired).

For decades, a human dentist has had to look at a panoramic X-ray (a wide, single photo of your whole jaw) and manually count these up. It's like a librarian trying to count every book on a shelf while wearing thick gloves; it takes time, and two different librarians might count slightly differently.

This paper introduces a digital assistant (an AI) that can look at that same X-ray and automatically write the report card for you. Here is how they built it and what they found, explained simply:

1. The Training: Teaching the Robot

The researchers taught the AI using a massive library of 705 X-rays (from the "DENTEX 2023" dataset). They didn't just show the AI pictures; they gave it a "answer key" where a human had already marked which teeth were healthy, which had cavities, and which were filled.

Think of this like a student studying for a final exam by reviewing 705 practice tests with the answers already written in red ink. The AI learned to spot the patterns of a "Healthy" tooth, a "Cavity," and a "Filled" tooth.

2. The Test: The Blind Challenge

To see if the AI was actually smart or just memorized the practice tests, the researchers gave it a brand new, secret test.

They took 100 X-rays from a different hospital that the AI had never seen before.
They asked two real, expert dentists to count the DMFT scores on these 100 images independently.
Then, they averaged the two dentists' answers to create a "Gold Standard" (the best possible human answer).
Finally, they let the AI take the same test.

3. The Results: How Did the AI Do?

The results were surprisingly good, but with one major catch.

The Overall Score: The AI's total DMFT score was almost as close to the "Gold Standard" human average as the two human dentists were to each other. If two humans agree 88% of the time, the AI agreed 90% of the time. In other words, the AI was as reliable as a second human dentist.
The "Missing" and "Filled" Wins: The AI was a superstar at spotting Missing teeth (empty spots) and Filled teeth (repairs). It was very accurate here, almost perfectly matching the humans.
The "Cavity" Struggle: The AI was much weaker at spotting Decayed teeth (cavities). It often missed small cavities or saw cavities where there weren't any.
- Why? The paper explains that panoramic X-rays are like looking at a city from a high airplane; you can see the big buildings (missing/filled teeth) clearly, but it's hard to see the small potholes (cavities) on the street level. Also, the AI was trained on data that was a bit too "generous" in calling things cavities, so it tends to over-count them.

4. The "Over-Counting" Quirk

Because the AI was so eager to find cavities, it tended to give a slightly higher DMFT score than the humans. On average, it added about 1.25 extra points to the score.

Imagine if you were grading a test and you accidentally gave every student 1 extra point because you were being too generous. The AI does this with cavities.
However, despite this over-estimation, 84% of the time, the AI's score was within just 3 points of the human experts' score.

5. Speed and Purpose

The AI is incredibly fast. It can analyze a full mouth X-ray in about 0.3 seconds (faster than a human can blink). A human dentist takes 1 to 5 minutes to do the same job.

What the paper actually says this AI is for:
The authors are careful to say this tool is not a replacement for a dentist. You still need a human to look at the patient. Instead, they see it as a helper tool (like a spell-checker for dentists) that:

Helps standardize the counting process so everyone gets the same result.
Reduces the workload for dentists by handling the easy parts (counting missing and filled teeth).
Could be used to quickly check the oral health of large groups of people (like a city-wide health survey) using old X-ray files.

The Bottom Line

This paper proves that an AI can look at a dental X-ray and count missing and filled teeth almost as well as a human dentist. It struggles a bit with spotting small cavities and tends to over-count them, but overall, it's a fast, reliable assistant that can help dentists and health organizations get a quick, standardized snapshot of oral health.

Technical Summary: Development and External Validation of an AI Pipeline for Automated DMFT Determination

Problem Statement
Untreated dental caries remains the most prevalent chronic condition globally, necessitating reliable epidemiological monitoring via the Decayed-Missing-Filled-Teeth (DMFT) index. While the DMFT index is the standard for World Health Organization-aligned oral health surveillance, its manual reconstruction from panoramic radiographs (OPGs) is time-consuming and subject to significant inter-rater variability. Current workflows rely on a combination of intra-oral inspection and imaging, but the single-acquisition OPG view is the only method to visualize the full dentition simultaneously. Although deep learning has advanced in adjacent tasks like caries detection and restoration identification, no system had previously been published that performs fully automated, patient-level DMFT determination from a single panoramic radiograph, validated against an external clinical cohort with rigorous uncertainty quantification.

Methodology
The study developed a single-stage deep-learning pipeline using a YOLO26x detector (Ultralytics), a native end-to-end object detection architecture.

Data and Annotation:
- Training: The model was trained on 705 annotated panoramic radiographs from the public DENTEX 2023 dataset. Annotations were re-coded from pathological labels into three target classes: Healthy, Caries (merging two caries grades), and Filled. Tooth-level restoration status was integrated via an external list mapped to FDI positions.
- External Validation: The system was tested on 100 anonymized OPGs from an independent university hospital (University Hospital Düsseldorf).
- Reference Standard: Two dentists independently annotated the 100 external images using WHO DMFT-only conventions (integer counts of D, M, F). The adjudicated reference was the component-wise mean of the two raters, rounded to the nearest integer.
Pipeline Architecture and Inference:
- Training: The model was fine-tuned using AdamW with a cosine-annealing learning rate schedule on 1024-pixel images with mosaic augmentation.
- Inference Strategy: To ensure robustness, the pipeline employed four-fold test-time augmentation (TTA) consisting of identity, horizontal flip, and two scale variants (0.83× and 1.20×). Predictions from these augmentations were fused using Weighted-Box-Fusion (WBF) at an IoU threshold of 0.55.
- Counting Logic:
  - Decayed (D) and Filled (F): Derived directly from detected bounding boxes for the respective classes.
  - Missing (M): Inferred rather than explicitly detected. The system calculated $M = \max(0, 28 - \text{count of present teeth})$ , adhering to a 28-tooth convention that excludes third molars due to frequent truncation or impaction in OPGs.
  - DMFT Index: Calculated as the sum $D + M + F$ .
Evaluation Metrics:
Performance was assessed using Intraclass Correlation Coefficient (ICC 2,1), Pearson correlation, Mean Absolute Error (MAE), and Bland–Altman analysis with 10,000-resample bootstrap confidence intervals. Agreement was benchmarked against the adjudicated human reference and compared to the inter-rater agreement between the two dentists.

Key Results

Overall Agreement: The AI pipeline achieved an ICC of 0.90 and a Pearson correlation of 0.91 against the adjudicated human reference. This performance was statistically indistinguishable from the inter-rater agreement of the two human dentists (ICC 0.88; difference +0.02, 95% CI -0.04 to 0.09).
Error Metrics: The Mean Absolute Error (MAE) for the AI was 2.21 DMFT units, comparable to the human inter-rater MAE of 2.13.
Bias and Limits of Agreement: Bland–Altman analysis revealed a systematic positive bias of +1.25 DMFT units (AI over-estimates), with 95% Limits of Agreement ranging from -4.0 to 6.5.
Component-wise Performance:
- Missing and Filled Teeth: Agreement was strong (Pearson $r = 0.90$ and $0.91$, respectively).
- Decayed Teeth: Agreement was weak (Pearson $r = 0.33$ ), identified as the primary source of error.
Cumulative Agreement: 84% of radiographs fell within ±3 DMFT units of the reference, and 90% within ±5 units.
Subgroup Analysis: Performance remained relatively uniform across image sharpness terciles and DMFT severity strata. The error did not decrease in caries-free mouths, confirming that the over-estimation is driven by systematic caries over-detection rather than case complexity.
Efficiency: Inference time was approximately 0.3 seconds per OPG on a consumer-grade NVIDIA RTX 3090 GPU, significantly faster than manual reading times.

Significance and Claims
The paper claims to present the first deep-learning system for the fully automated determination of the DMFT index from panoramic radiographs, validated on an independent external cohort. Its primary significance lies in demonstrating that an AI pipeline can achieve overall agreement with trained dentists that is comparable to the agreement between the dentists themselves.

The authors position the system not as a replacement for clinical examination, but as a decision-support and documentation tool. They highlight that the pipeline is particularly robust for assessing missing and filled teeth—components critical for treatment planning and population surveillance—while acknowledging that the decayed component remains a limitation due to the inherent resolution constraints of panoramic radiography for caries detection. The study underscores the feasibility of using routinely acquired OPGs for standardized, scalable oral health surveillance, potentially reducing examiner workload and inter-rater variability in epidemiological studies.