Obtaining full-arch implant scan with smartphone video and deep learning: An in vitro investigation on trueness and precision.

Abstract

Purpose: To investigate the accuracy of complete-arch implant scans generated by a smartphone camera and a deep learning model.

Materials and methods: A deep learning model was trained to generate 3D scans from smartphone videos using a maxillary edentulous model with 6 implants and scan bodies (SBs). Three test groups were created: (1) deep learning 3D reconstruction with 1500 training epochs (DL1), (2) deep learning 3D reconstruction with 5000 training epochs (DL2), and (3) scans obtained from an intraoral scanner (IOS). Each method was repeated 10 times, with a desktop scanner scan as the reference. Test scans were aligned to the reference using two methods: (a) aligning all SBs to evaluate the overall fit, and (b) aligning just the first and second SBs to simulate passive fitting test of multiple implant-supported prostheses. Linear deviations from the reference model (trueness) and within each group (precision) were analyzed.

Results: For the overall fit, the DL2 group (67.69 ± 33.29 µm) showed significantly better (p < 0.05) mean trueness than the DL1 group (127.82 ± 73.07 µm), and similar trueness to the IOS group (57.42 ± 36.09 µm). However, the DL2 group (98.12 ± 59.85 µm) showed worse (p < 0.05) precision compared to the IOS group (64.54 ± 42.53 µm). In the virtual passive-fitting test, the DL2 group showed similar trueness and accuracy compared to the IOS group.

Conclusions: In the in vitro environment, combining smartphone videos with a deep learning model generated full arch implant scans with accuracy similar to an IOS. Although this accuracy is not good enough for clinical application, this approach shows promise as a potential direction for future development in economical full-arch implant scanning.