The research, led by Orkun Elçi and Mücahid Yıldırım of the Department of Orthodontics at Necmettin Erbakan University in Konya, Turkey, together with colleagues in dentomaxillofacial surgery and radiology, set out to build and rigorously test a deep learning system capable of grading midpalatal suture maturation automatically from cone-beam computed tomography scans, the low-dose three-dimensional imaging modality that has become a staple of modern orthodontic offices. Their tool of choice was YOLOv11x, the largest configuration of the latest generation of the You Only Look Once family of object detection networks, an architecture prized for its ability to locate and classify objects within an image in a single forward pass rather than in the slower multi-stage fashion of older detectors.

To train the model, the team assembled a strikingly large dataset by the standards of dental imaging research: 1,501 CBCT examinations acquired between January 2021 and January 2025 from patients aged 10 to 25 years, precisely the window during which the midpalatal suture transitions from open to fused. Each examination was graded according to a modified version of the Angelieri classification, the most widely used staging system for suture maturation, and then collapsed into three clinically meaningful groups labeled AB, C, and DE. The grouping is not arbitrary. Stages A and B represent a suture that is still relatively open and responsive to expansion, stage C marks an intermediate state of partial fusion, and stages D and E indicate a suture that has largely or completely ossified, where conventional expansion is likely to fail.

From the volumetric scans, the researchers extracted standardized axial slices, the cross-sectional views in which the suture’s maturation signature is most visible, and used these two-dimensional images as the input for the network. A total of 1,200 images were used for training and 150 for validation, all drawn from a single scanner manufacturer, Planmeca, so that the model could first learn the task under consistent imaging conditions. The crucial test, however, came afterward. The team held out 151 CBCT images acquired with an entirely different scanner, a J. Morita system, and asked the trained model to classify sutures it had never seen, produced by hardware it had never encountered. This cross-device evaluation matters enormously, because an algorithm that only works on the brand of machine it was trained on is of little use in the real world, where clinics deploy a patchwork of imaging systems.

The results were encouraging on every metric the researchers reported. On the held-out Morita test set, the YOLOv11x model achieved a mean average precision at the standard 0.5 intersection-over-union threshold of 0.910, and a stricter mean average precision averaged across thresholds from 0.5 to 0.95 of 0.619. Broken down by class, the average precision at 0.5 reached 0.943 for the AB group, 0.838 for the intermediate C group, and 0.949 for the fused DE group. In plain terms, the network was excellent at both finding the suture region on the image and assigning it to the correct maturity category, with the middle stage proving the hardest, as it often does for human readers as well.

The confusion matrix analysis, which tallies how often the model’s answers matched the reference labels, painted a similarly consistent picture. Precision values came in at 0.818 for AB, 0.708 for C, and 0.817 for DE, while recall values were 0.865, 0.810, and 0.860 respectively, yielding F1-scores of 0.841, 0.756, and 0.838. Of the 151 test images, 128 were classified correctly, an overall accuracy of 84.8 percent across a scanner the model had never seen during training. For a task that hinges on detecting subtle bridging bone within a thin fibrous joint, and that must generalize across different reconstruction algorithms and dose settings, that level of performance marks a meaningful step forward.

Why does this matter clinically? Rapid maxillary expansion is one of the most common interventions in growing patients, used to correct crossbites, relieve crowding, and widen a constricted airway. But the success of the procedure depends heavily on timing. Studies of the Angelieri stages have repeatedly shown that expansion outcomes differ sharply between an open suture and a fused one, and that clinicians reading the same scan can disagree about which stage they are looking at. Cone-beam CT offers far more detail than two-dimensional films, yet it also introduces variability in slice selection and interpretation. An algorithm that returns the same answer every time, regardless of operator fatigue or experience, could reduce that inter-observer noise and give treatment planners a reproducible second opinion at the moment of decision.

The study’s design also reflects a growing sophistication in how medical AI is evaluated. Rather than reporting only accuracy on a random split of same-scanner images, the Turkish team deliberately separated training data from a device-specific test set, quantified localization quality with mean average precision, and examined per-class errors through the confusion matrix. This kind of reporting makes it possible to see precisely where the model struggles, in this case the intermediate C stage, where partial bridging produces ambiguous visual patterns, and where it excels. The authors note that the model showed promising localization and classification performance across clinically relevant maturation stages on the selected axial images, a carefully worded conclusion that acknowledges both the strength of the results and the constrained scope of the evaluation.

Limitations remain before such a system could enter routine practice. The model was trained and tested on standardized axial slices rather than full volumes, meaning that a deployed tool would need reliable slice selection, a problem the detection framework itself may eventually absorb. The dataset came from a single institution, and although the cross-device test on J. Morita hardware is a strong signal of generalizability, broader multi-center validation across more scanners, age distributions, and anatomical variations would be the natural next step. The retrospective design, ethical approval from the Research Ethics Committee of the Faculty of Dentistry at Necmettin Erbakan University, and anonymous processing of all data are appropriate for a proof-of-concept, but prospective studies measuring whether AI-assisted staging actually changes treatment decisions and outcomes would be needed to establish clinical value.

Still, the trajectory is hard to ignore. Orthodontics has been slower to adopt diagnostic AI than radiology or pathology, partly because musculoskeletal and dental structures resist the clean, high-contrast boundaries that make tasks like lung nodule detection tractable. The midpalatal suture, a structure a few millimeters wide whose maturity is encoded in the texture and continuity of thin bone bridges, is exactly the kind of subtle target where modern detection networks, given enough well-labeled data, can begin to match expert judgment. With 1,501 scans, a state-of-the-art YOLOv11x backbone, and a genuinely independent cross-device test, this study offers one of the more convincing demonstrations to date that the suture’s developmental clock can be read automatically. If subsequent multi-center trials confirm these numbers, the humble palatal seam may become one of the latest additions to the growing list of anatomical structures that no longer depend solely on the human eye for their interpretation.

Subject of Research: AI-based classification of midpalatal suture maturation stages from cone-beam computed tomography images for orthodontic treatment planning

Article Title: AI–assisted evaluation of midpalatal suture maturation using CBCT: a YOLOv11 study

Article References: Elçi, O., Yıldırım, M., Üstün, M., & Altındağ, A. (2026). AI–assisted evaluation of midpalatal suture maturation using CBCT: a YOLOv11 study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02886-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02886-z

Keywords: artificial intelligence, deep learning, YOLOv11, cone-beam computed tomography, midpalatal suture, orthodontics, maxillary expansion, Angelieri classification, medical imaging, object detection, BMC Medical Imaging, diagnostic automation