Answers are grounded in this paper and cite their sources.
To address the difficulty of quantitatively assessing the safety state during robot-assisted ultrasound-guided thyroid nodule puncture, this study proposes an in-process risk-grading method for experimentally defined process-risk states integrating force sensing, end-effector Z-position, and ultrasound imaging information. A phantom experimental platform was constructed to acquire force/torque signals, end-effector Z-position data, and ultrasound images during needle insertion, with window-level samples used as the basic analytical unit. Based on six observable risk factors-abnormal axial insertion force, abnormal lateral force, abnormal torque, loss of needle visibility, loss of nodule visibility, and blurring of ultrasound monitoring images-four-level in-process risk labels were constructed according to single-factor abnormality severity and multi-factor accumulation rules. Multimodal features derived from force sensing, end-effector Z-position, and ultrasound imaging were then extracted to develop an XGBoost-based four-level risk-grading model. A total of 536 valid window-level samples were obtained from 100 independent puncture trials. The results of 20 repetitions of five-fold cross-validation grouped by experiment ID showed that the model achieved an accuracy of 0.7318 ± 0.0489, a Macro-F1 score of 0.7005 ± 0.0755, a Macro-OVR AUC of 0.8998 ± 0.0311, and a quadratic weighted kappa (QWK) of 0.7293 ± 0.0975. The severe underestimation rate was 0.0272 ± 0.0232, and the Level IV recall was 0.8080. Algorithm comparison, modality ablation, and model interpretation analyses showed that the proposed method effectively integrated ultrasound visual-monitoring status, insertion-stage information represented by end-effector Z-position, and mechanical interaction information for window-level in-process risk discrimination. These results provide preliminary evidence for the feasibility of multimodal machine learning-based in-process risk grading for robot-assisted ultrasound-guided puncture under phantom conditions and provide a methodological basis for future research on online risk warning and in-process safety-state perception.
No takes yet. Share an insight, caveat, or question.
郭雅佩 et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: