YOLO11-guided swin transformer for molar occlusion classification

Area of Science:

• Dentistry and Dental Technology
• Artificial Intelligence in Medicine
• Computer Vision

Background:

• Molar occlusion identification is crucial for orthodontic diagnosis and treatment planning.
• Manual assessment of intraoral photographs is time-consuming and prone to variability.
• Current deep learning (DL) methods lack precise anatomical localization or rely on manual regions of interest, limiting clinical use.

Purpose of the Study:

• To propose a novel DL framework for automated molar occlusion classification.
• To integrate YOLOv11 for anatomical localization and Swin Transformer (Swin-T) for occlusion recognition.
• To enhance the accuracy and clinical applicability of automated orthodontic assessments.

Main Methods:

• A DL framework combining YOLOv11 for molar region instance segmentation and Swin-T for occlusion classification was developed.
• YOLOv11 automatically localized molars, and Swin-T analyzed prediction-guided crops using hierarchical shifted-window attention.
• A dataset of 1101 intraoral images across five occlusion classes was utilized for training and validation.

Main Results:

• The localization stage achieved a high Dice coefficient of 96.82%.
• The framework attained a classification accuracy of 98.17%, surpassing expert orthodontists.
• The model demonstrated superior performance compared to ResNet50V2, DenseNet201, MobileNetV3-Large, EfficientNetV2-S, and ViT-B/16, with perfect recall for specific classes.

Conclusions:

• The proposed DL framework offers a highly accurate and automated solution for molar occlusion classification.
• The system's ability to focus on clinically relevant structures validates its anatomical plausibility.
• This technology provides a strong foundation for integrating DL-driven decision support into digital orthodontic workflows.