Hybrid Attention-CNN and Vision Transformer Models Compared for CBCT Classification of Third-Second Molar Interactions

Transformer-based and attention-CNN deep learning models outperformed conventional CNNs for classifying third-second molar relationships on CBCT; an ensemble achieved the highest macro-AUC. Differentiating contact from independent cases remained most challenging.

• Deep learning assessment of 306 third-second molar units from 162 CBCT scans categorized interactions as independent, contact, or resorption.
• Attention-based and Vision Transformer models improved classification over conventional CNNs, with ensemble learning further enhancing reliability.
• Distinguishing contact from independent cases was the most difficult task, while resorption was identified more consistently.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.