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

Area of Science:

• Computer Vision
• Medical AI
• Dental Technology

Background:

• Dynamic navigation and robot-assisted implant surgery require precise intraoral perception.
• Marker-based tracking is complex and prone to occlusions, necessitating markerless solutions.
• Current markerless methods often lack the speed or accuracy needed for real-time dental workflows.

Purpose of the Study:

• To evaluate a single-stage YOLO instance segmentation model (YOLO-seg) as a markerless perception layer for dental navigation.
• To assess YOLO-seg's ability to provide accurate per-tooth delineation and low inference latency.
• To compare YOLO-seg's performance against a two-stage YOLO + SAM baseline model.

Main Methods:

• Trained YOLO-seg on a diverse intraoral RGB image corpus (400 training, 20 validation, 100 testing images).
• Evaluated segmentation quality on a 50-image clinical test set using per-instance matching, per-class union, and global union metrics.
• Assessed runtime performance under matched inference-only and end-to-end conditions on 100 frames at 640x640 resolution using an NVIDIA RTX A2000 GPU.

Main Results:

• YOLO-seg significantly outperformed the YOLO + SAM baseline across all segmentation metrics (e.g., AP50 = 0.716 vs 0.383, recall = 0.973 vs 0.398).
• YOLO-seg achieved substantially lower inference latency: 27.08 ms/frame (36.9 FPS) compared to 1302.78 ms/frame (0.77 FPS) for the baseline.
• The single-stage YOLO-seg model demonstrated a ~48-fold speed advantage in inference time.

Conclusions:

• YOLO-seg shows significant potential as a 2D perception module for real-time dental navigation systems.
• The model effectively balances high segmentation fidelity with predictable, low inference latency.
• Results support YOLO-seg's integration as a front-end for 3D reconstruction and robot-assisted dental procedures.