Table 4.

The evaluation results of core AI tools.

AI tool category	Representative examples	Accuracy	Generalizability	Safety	Clinical integration	Regulatory status	References
Segmentation models	nnUNet, U-Net-based rotator cuff model	DSC: 0.86–0.95 (humerus/glenoid/rotator cuff)	Limited by training data (few obese patients)	No direct procedural risk; data privacy dependent on hospital systems	Compatible with ultrasound/MRI; requires radiologist collaboration	Not independently regulated (integrated into software)	Mu et al. (2021), Dai et al. (2024), Medina et al. (2021), and Alipour et al. (2024)
Navigation systems	ExactechGPS®, Joint VTS	Injection accuracy: 90 ~ 96.6%	Moderate (validated in RSA/TSA; limited in frozen shoulder)	Complication rate < 1%; real-time vibration alerts	Integrates with ultrasound/robots; 1–2 weeks learning curve	NMPA Class III/FDA 510(k)/CE Class IIb	Xu et al. (2025), Andriollo et al. (2024), and Kuratani et al. (2022)
Real-time learning platforms	Federated averaging-based closed-loop systems	Adaptive accuracy improvement: 5 ~ 10% after 100 cases#	High (multi-center data integration)	Edge computing protects privacy; model updates require validation	validationSeamless with intraoperative workflow; no additional operator burden	Regulatory gap (continuous learning not fully standardized)	Xu et al. (2025) and Schweihoff et al. (2021)
Specialized tools for subgroups	AI models for BMI > 35 patients	First-pass success rate: 85% (vs. 65% for conventional AI)	High (targets obese/large tear patients)	Reduces soft tissue injury risk by 20%	Requires high-resolution ultrasound; short learning curve	NMPA/FDA pending (pilot stage)	Huang et al. (2015) and Wu et al. (2025)

The performance of real-time learning platforms (5 ~ 10% accuracy improvement after 100 cases) is inferred from federated learning technical characteristics, not direct clinical trial data.