Cascade CNN detects dental landmarks on 3D models for automated orthodontic measurements

A projection-based cascade CNN detected 34 dental landmarks on 3D digital models with mean error 0.66 mm; arch length discrepancy (ALD) error was 0.90 mm, comparable to intraobserver error (0.95 mm). Processing time <4 s per model. External validation ALD error: 1.70 mm.

• Tooth region detection success rate was 99.44% with mean IoU 0.91.
• AI-calculated ALD error (0.90 mm) did not differ significantly from intraobserver error (0.95 mm).
• Preliminary external validation on 22 models from another institution showed a higher ALD error (1.70 mm), indicating need for larger-scale validation.

Automated summary

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