Multiteacher Knowledge Distillation Improves Canine Scoring on Panoramic Radiographs

A student model using multiteacher knowledge distillation outperformed individual teacher models for classifying impacted maxillary canines on panoramic radiographs, achieving 80.43% accuracy and AUC 0.89 with differential teacher weighting on a hold-out test set of 92 images.

• The student model with equal teacher weighting achieved 79.35% accuracy and an AUC of 0.88.
• Differential teacher weighting further improved performance to 80.43% accuracy and AUC 0.89.
• The approach demonstrates that multiteacher knowledge distillation with complementary preprocessing variants can be effective for radiographic scoring with limited annotated data.

Automated summary

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