Deep Learning for Automatic Segmentation of Bone Sequestra on Orthopantomographs: A UNet-Based Approach
Key facts
- Who runs it
- Izmir Katip Celebi University
- Where
- Türkiye1 site in 1 countryChoose a country to see its sites
- Trial code
- Status
- Completed
- People taking part
- 120 people enrolled
- Started
- 1 Mar 2024
- Ended
- 1 Mar 2025
- Type of study
- Observational
- Treatments tested
- Radiographic Evaluation
- Record updated
- 9 September 2026
Who can take part
- Age
- 18 years to 99 years
- Sex
- Women and men
- Healthy volunteers
- Cannot take part
The registry criteria, word for word
Show all 6 criteriaShow less
Who can take part
- Patient with exposed bone over 8 weeks with a history of bisphosphonates or antiangiogenic drugs,
- Patient with a history of previous radiotherapy,
- Patients with necrotic exposed bone with or without a history of trauma
- Patients with osteomyelitis
Who cannot take part
- Patients with no clear vision of the borders of the sequestra,
- Panoramic images with artifacts in and around the lesion
This trial has no sites in United States.
It has sites in one country. Choose a country to see its sites.
About the trial
What the trial wants to find out
This study developed and evaluated a UNet-based deep learning model for automatic segmentation of bone sequestra on orthopantomographs (OPGs), with the goal of improving diagnostic efficiency and reducing inter-observer variability in osteomyelitis, osteoradionecrosis, and medication-related osteonecrosis of the jaw. A total of 120 anonymized OPGs with 134 annotated sequestra were used for training. Images were preprocessed with min-max normalization and augmented to enhance robustness. Manual expert annotations served as ground truth. A UNet architecture, optimized with Dice loss and the AdamW optimizer, was trained for 300 epochs. Performance was assessed using Dice coefficient, Intersection over Union (IoU), precision, recall, F1-score, and accuracy. ROC analysis and confusion matrix evaluations were performed. Agreement with clinicians was quantified using the Intraclass Correlation Coefficient (ICC). The model achieved a best-checkpoint validation Dice of 0.79 and IoU of 0.93. On the test set, performance included a Dice of 0.79, IoU of 0.74, recall of 0.81, precision of 0.81, and F1-score of 0.81. ROC analysis showed balanced discriminative performance with an AUC of 0.76. The confusion matrix indicated strong lesion/background classification, with minor under-segmentation. ICC analysis showed excellent agreement with an experienced surgeon (ICC = 0.85), though lower with a less experienced dentist (ICC = 0.57). The proposed UNet-based model enables efficient segmentation of sequestra, reducing annotation time by >90% while maintaining diagnostic accuracy. The model highlights the clinical utility of AI-assisted decision support in maxillofacial radiology.
Treatments tested
- Radiographic Evaluation
Data and sources
- Registry record
- Listed on ClinicalTrials.gov.How to verify a trial in the registry
- Record updated
- 9 September 2026