Real-Time Markerless Tooth Detection Towards Dynamic Robot-Assisted Dental Implant Navigation.

Background/

Objectives

Dynamic navigation and robot-assisted implant workflows depend on robust intraoral perception. Marker-based tracking introduces workflow complexity and is sensitive to occlusions, motivating markerless alternatives. This study evaluates whether a single-stage YOLO instance segmentation model (YOLO-seg) can provide a practical markerless perception layer for dental navigation, combining accurate per-tooth delineation with low, predictable inference latency.

Methods

YOLO-seg was trained end to end on an intraoral RGB corpus of 400 training, 20 validation, and 100 testing images, combining a public source and a partner-hospital in-house set. A two-stage YOLO + SAM baseline was implemented for comparison. Segmentation quality was evaluated on a 50-image held-out clinical test set at three complementary levels (per-instance matching, per-class union, and global union), with paired Wilcoxon signed-rank tests, Cliff's delta effect sizes, and 95% bootstrap confidence intervals. Runtime was assessed under matched inference-only and end-to-end conditions on N = 100 frames at a 640 × 640 resolution on an NVIDIA RTX A2000 GPU.

Results

YOLO-seg significantly outperformed YOLO + SAM across all primary metrics, with very large effect sizes (Cliff's delta: 0.76-0.94; Wilcoxon p < 10-8 on every metric except precision at IoU ≥ 0.5). YOLO-seg reached AP50 = 0.716 and recall = 0.973 versus 0.383 and 0.398 for YOLO + SAM. Under matched inference-only timing, YOLO-seg ran at 27.08 ms per frame (36.9 FPS) versus 1302.78 ms (0.77 FPS), an approximately 48-fold latency gap intrinsic to the two-stage forward pass.

Conclusions

YOLO-seg shows strong potential as a 2D perception module for dental navigation, balancing per-instance segmentation fidelity with real-time feasibility under the tested conditions. These results support its use as a 2D perception front-end for future integration with stereo-based 3D reconstruction and robot-assisted navigation; 3D registration accuracy, implant-placement error, and robotic execution remain outside the scope of the present study.