AI pipeline generates dental crown proposals from partial scans in 3 minutes

A new classification-guided restoration assistance pipeline developed by Rabin Kunwar and 13 colleagues offers a practical alternative to end-to-end generative methods for single-unit crown restoration. The system accepts a raw intraoral scan and a target FDI tooth number, then produces an initial crown proposal for clinician refinement. It addresses a key gap: most segmentation networks trained on full-arch scans fail on partial scans, while generative methods often over-smooth occlusal details.

The pipeline operates in three phases: data preparation and pose standardization, scan-type-routed segmentation, and context-aware crown proposal generation using retrieval and Blender-based fitting. A DGCNN classifier categorizes the scan into one of five anatomical types, followed by coarse-to-fine RANSAC+ICP registration to standardize the jaw coordinate frame, then graph-cut optimization refines tooth-gingival boundaries. On 1,958 partial scans, it achieved macro-average DSC 0.9249, recall 0.8919, and precision 0.9615 across 17 semantic classes. For full arches, a fine-tuned model reached DSC 0.9347. The target tooth and its neighbors achieved DSC 0.9468–0.9569 with sub-millimeter centroid errors (0.2666–0.2774 mm). The retrieval module uses DGCNN embeddings and cosine similarity over neighboring and opposing teeth, followed by spline-guided alignment and Blender Python API refinement. The entire process completes in 2.5–3.5 minutes, offering a fast, accurate starting point for clinicians.