Cephalometric landmark annotation using transfer learning: Detectron2 and YOLOv8 baselines on a diverse cephalometric...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthodontics

Background:

• Cephalometric landmark annotation is crucial for orthodontic diagnosis and treatment planning.
• Manual annotation is time-consuming and subjective, necessitating automated solutions.
• Transfer learning offers a potential approach to improve accuracy in radiographic analysis.

Purpose of the Study:

• To evaluate advanced transfer learning techniques for accurate cephalometric landmark annotation.
• To compare the performance of Detectron2 and YOLOv8 frameworks on cephalometric images.
• To assess the impact of dataset size on model performance.

Main Methods:

• Utilized state-of-the-art pose estimation models: Detectron2 (ResNet backbones) and YOLOv8 (various sizes).
• Trained and evaluated models on the DiverseCEPH19 dataset (1692 images, 19 landmarks) and a benchmark dataset (400 images).
• Analyzed model performance across different image categories and dataset sizes.

Main Results:

• Detectron2 demonstrated superior performance, achieving 85.89% accuracy on the DiverseCEPH19 dataset.
• YOLOv8 achieved 72.92% accuracy on the same dataset.
• Detectron2 consistently outperformed YOLOv8 on the smaller benchmark dataset.

Conclusions:

• Detectron2 shows suitability for high-precision cephalometric annotation applications.
• While Detectron2 excels in accuracy, YOLOv8 baselines may be favored when considering model size and inference time.
• Further research can optimize model selection based on specific application requirements.