Compact Involutional Transformer for Automated Detection of Pediatric Tooth Number Anomalies on Panoramic Radiographs
摘要
Pediatric tooth number anomalies can compromise occlusion, craniofacial development, and long-term treatment planning. This study introduces the Compact Involutional Transformer (CIT), a novel transformer architecture for automated detection of permanent tooth germ deficiency and supernumerary teeth on pediatric panoramic radiographs. The model is a transformer fed by an adaptive involution-based tokenizer, combining locality-aware tokenization with contextual self-attention in a compact design. Pediatric panoramic radiographs (n = 1170) were retrospectively collected from patients with completed diagnoses and treatments, and radiographic labels were curated and verified by an experienced pediatric dentist. Performance was evaluated on both multi-class (germ deficiency, normal, and supernumerary teeth) and binary tasks. In the three-class setting, CIT outperformed state-of-the-art baselines, achieving 96.00% accuracy, 95.29% F1, 95.76% ROC-AUC, and 93.28% Matthews correlation coefficient. The decision process was examined using Grad-CAM visualizations. Model predictions were benchmarked against two independent dentist cohorts with different experience levels; Bonferroni adjusted McNemar’s tests (m = 5;