Applications of artificial intelligence in orthodonticsa bibliometric and visual analysis

Alessandro Polizzi, Mattia Boato, Sara Serra, Vincenzo D’Antò, Rosalia Leonardi•View original
HighlightsAccessiblealloy voice
Imagine your orthodontist planning your treatment like a pilot plotting a route. Faster, safer, and with fewer surprises. That’s the promise of artificial intelligence in orthodontics, and Polizzi and colleagues set out to see where that promise is actually taking shape by mapping the entire research landscape. They didn’t test one algorithm; they took a panoramic view. Using the Web of Science, they scanned papers from 1985 to 2024, pulled 912 records, and, after screening, kept 381 that truly tied artificial intelligence to orthodontics. Think of it as a census of ideas. What does "mapping" mean here? It’s bibliometrics — turning the world of papers, authors, and keywords into networks to spot who’s collaborating with whom, which topics are surging, and where the gaps are. The team used tools called CiteSpace and VOSviewer to build these maps. Two independent reviewers agreed on which studies to include to a near-perfect degree, achieving a Cohen’s kappa of 0.91, and the structure of the map held up well: the clusters were cleanly separated and consistent, with a modularity of 0.4212 and a silhouette of 0.7708. In plain terms, the patterns they observed are likely real, not random noise. The headline trend is speed. About 83 percent of all the papers appeared in just the last three and a half years, and 2023 alone was the peak, with 116 articles. That’s a field hitting the gas. It’s also a global surge, but not evenly spread. Roughly two-thirds of the work comes from China, the United States, and South Korea, with China not just prolific but well-connected in the collaboration web, showing strong network influence, reflected by a centrality of 0.19. So the centers of gravity are clear, and they’re talking to each other more than ever. Where is this work showing up? Not just in orthodontic journals, but also hand-in-hand with computer science and medical imaging. The American Journal of Orthodontics and Dentofacial Orthopedics, The Angle Orthodontist, and Lecture Notes in Computer Science rank among the most cited venues, underscoring that cross-disciplinary pull. On the institutional side, names like Seoul National University, Yonsei, and Peking University appear repeatedly, and the network hubs are telling: Peking University and Hyung Hee University sit at the top for influence in the map, with centralities of 0.30 and 0.23. That’s where collaborations and citations tend to converge. Drill down to what the algorithms actually do, and two jobs dominate. First, cephalometric landmarking — having software pinpoint key anatomical points on head X-rays. Second, image segmentation — drawing clean borders around teeth, bone, and soft tissue in scans. The keywords back this up. "Artificial intelligence" appears 137 times, while momentum waves — those bursts of sudden activity — crest around segmentation earlier, peaking near 2017, and then around automated landmark detection in 2020 to 2021. Ten thematic clusters emerge from the network, with labels like tooth segmentation, cephalometric diagnosis, and image segmentation, painting a field that’s broad but coherently organized. So, is it clinic-ready? Not quite across the board. Accuracy varies. Reviews note that in two-dimensional X-rays, these systems are less likely to land within a 2 millimeter window than in three-dimensional scans, and even that 2 millimeter target isn’t always good enough for every clinical decision. Models that suggest tooth extractions perform well on paper, but the evidence is mixed across datasets. That’s why Polizzi’s team keeps coming back to rigor: larger and more diverse data, multi-center collaborations, and consistent reporting. They point clinicians and researchers to evolving playbooks — TRIPOD-AI and CLAIM for transparent reporting, CONSORT-AI and DECIDE-AI for trials and early clinical evaluation, PROBAST-AI and CLEAR for judging bias and evidence quality, and the FUTURE-AI framework to keep clinical needs front and center. Here’s the takeaway to carry with you. The field is exploding, anchored by a few core tasks that matter for real patients, and it’s increasingly connected across countries and disciplines. The map looks solid, and the momentum is real. Now the work is to turn that speed into reliability, so the pilot’s route isn’t just fast, it’s safe — every time.

Imagine your orthodontist planning your treatment like a pilot plotting a route. Faster, safer, and with fewer surprises. That’s the promise of artificial intelligence in orthodontics, and Polizzi and colleagues set out to see where that promise is actually taking shape by mapping the entire research landscape.

They didn’t test one algorithm; they took a panoramic view. Using the Web of Science, they scanned papers from 1985 to 2024, pulled 912 records, and, after screening, kept 381 that truly tied artificial intelligence to orthodontics. Think of it as a census of ideas.

What does "mapping" mean here? It’s bibliometrics — turning the world of papers, authors, and keywords into networks to spot who’s collaborating with whom, which topics are surging, and where the gaps are. The team used tools called CiteSpace and VOSviewer to build these maps.

Two independent reviewers agreed on which studies to include to a near-perfect degree, achieving a Cohen’s kappa of 0.91, and the structure of the map held up well: the clusters were cleanly separated and consistent, with a modularity of 0.4212 and a silhouette of 0.7708. In plain terms, the patterns they observed are likely real, not random noise.

The headline trend is speed. About 83 percent of all the papers appeared in just the last three and a half years, and 2023 alone was the peak, with 116 articles. That’s a field hitting the gas.

It’s also a global surge, but not evenly spread. Roughly two-thirds of the work comes from China, the United States, and South Korea, with China not just prolific but well-connected in the collaboration web, showing strong network influence, reflected by a centrality of 0.19. So the centers of gravity are clear, and they’re talking to each other more than ever.

Where is this work showing up? Not just in orthodontic journals, but also hand-in-hand with computer science and medical imaging. The American Journal of Orthodontics and Dentofacial Orthopedics, The Angle Orthodontist, and Lecture Notes in Computer Science rank among the most cited venues, underscoring that cross-disciplinary pull.

On the institutional side, names like Seoul National University, Yonsei, and Peking University appear repeatedly, and the network hubs are telling: Peking University and Hyung Hee University sit at the top for influence in the map, with centralities of 0.30 and 0.23. That’s where collaborations and citations tend to converge.

Drill down to what the algorithms actually do, and two jobs dominate. First, cephalometric landmarking — having software pinpoint key anatomical points on head X-rays. Second, image segmentation — drawing clean borders around teeth, bone, and soft tissue in scans.

The keywords back this up. "Artificial intelligence" appears 137 times, while momentum waves — those bursts of sudden activity — crest around segmentation earlier, peaking near 2017, and then around automated landmark detection in 2020 to 2021. Ten thematic clusters emerge from the network, with labels like tooth segmentation, cephalometric diagnosis, and image segmentation, painting a field that’s broad but coherently organized.

So, is it clinic-ready? Not quite across the board. Accuracy varies.

Reviews note that in two-dimensional X-rays, these systems are less likely to land within a 2 millimeter window than in three-dimensional scans, and even that 2 millimeter target isn’t always good enough for every clinical decision. Models that suggest tooth extractions perform well on paper, but the evidence is mixed across datasets. That’s why Polizzi’s team keeps coming back to rigor: larger and more diverse data, multi-center collaborations, and consistent reporting.

They point clinicians and researchers to evolving playbooks — TRIPOD-AI and CLAIM for transparent reporting, CONSORT-AI and DECIDE-AI for trials and early clinical evaluation, PROBAST-AI and CLEAR for judging bias and evidence quality, and the FUTURE-AI framework to keep clinical needs front and center.

Here’s the takeaway to carry with you. The field is exploding, anchored by a few core tasks that matter for real patients, and it’s increasingly connected across countries and disciplines. The map looks solid, and the momentum is real.

Now the work is to turn that speed into reliability, so the pilot’s route isn’t just fast, it’s safe — every time.