How AI and 3D Scans Are Transforming Clear Aligner Planning

/in /by

At Tristar, we’re committed to combining product excellence with rigorous research. Today we’re excited to share the publication of our paper describing LEAP — a deep learning system that classifies malocclusions from 3D intraoral scans and integrates directly into orthodontic CAD workflows to support clear aligner planning.

What LEAP does
LEAP converts standard tessellation language (STL) intraoral scans into voxel grids and processes them with a modified EfficientNet-3D convolutional neural network. The system outputs ranked predictions with confidence scores and applies hierarchical rules to produce clinically meaningful main and subclass combinations (for example, Class II Division 1 with open bite). Integration into CAD software enables automated classification at the point of care, helping clinicians move from scan to treatment design more efficiently.

Key results
The LEAP system demonstrated strong performance across eight clinically relevant malocclusion categories. “The LEAP system achieved robust classification performance, with 90.8% accuracy, 87.5% precision, 91.3% recall, and an F1 score of 89.2% across the eight malocclusion classes.” The model performed particularly well on anterior open bite cases (F1 ≈ 94.7%), while categories with greater morphological variability, such as Class II Division 2 and severe crowding, showed lower but still clinically useful performance (F1 ≈ 83%).

Clinical implications

  • Speed: Automated classification reduces time spent on initial diagnosis and case triage.
  • Consistency: Standardized outputs reduce interclinician variability and support quality control.
  • Integration: Delivering results inside CAD software streamlines the transition from diagnosis to aligner design.
  • Caveat: For complex or ambiguous cases, clinician oversight remains essential; LEAP is a decision‑support tool, not a replacement for clinical judgment.

Technical highlights

  • Dataset: 841 anonymized STL scans labeled and cross-validated by experienced orthodontists.
  • Preprocessing: Mesh voxelization at 128³/256³ resolutions preserves 3D geometry for volumetric learning.
  • Model: EfficientNet-3D adapted for single-channel voxel input; GPU-accelerated inference supports real-time use.
  • Logic: Hierarchical classification logic separates primary sagittal relationships from secondary occlusal traits to produce interpretable outputs.

Next steps and collaboration
We plan to expand dataset diversity, validate LEAP across additional scanners and clinical sites, and explore extensions that provide treatment recommendations and automated staging for clear aligner workflows. We welcome collaboration with clinics, universities, and industry partners interested in multicenter validation or pilot integration.

Read the paper
For full technical details, performance tables, and methodology, please read the published paper here.

#TristarLEAP #ClearAligners #3DScan #DentalAI