AI-assisted recurrent laryngeal nerve identification during endoscopic/robotic thyroid surgery based on the CMC-UNet model: a multicenter retrospective study.

During endoscopic or robotic-assisted thyroid surgery, the field of view may be restricted by tissue swelling or bleeding. These Limitations make delicate surgical manipulation in the confined space more challenging. This study proposes an artificial intelligence model designed to identify the recurrent laryngeal nerve. During thyroid surgery and thereby reduce the risk of accidental injury. This study retrospectively collected imaging data from three tertiary care hospitals, comprising 7482 images from 103 patients. The model relied on this dataset to optimize training. A variety of improvement strategies were integrated into the CMC-UNet model to enhance the accuracy and robustness of recurrent laryngeal nerve recognition. The Dice coefficient, Intersection over Union (IoU), and other metrics were used to evaluate the performance of the model in identifying the RLN. The Dice coefficient reached 0.8575, while the IoU achieved 0.7506. Compared with alternative models, the proposed model demonstrates a clear advantage in accuracy. Results from the controlled experiment indicate that anatomical variations in the course of the recurrent laryngeal nerve did not reduce the robustness of the model. Independent expert evaluations and inter-group experiments further confirmed the reliability and clinical applicability of the model, although variations may still occur across different surgical contexts. The CMC-UNet model achieved high accuracy in identifying the recurrent laryngeal nerve (RLN), with performance comparable to that of senior medical experts. The results of assisted identification may support clinical decision-making by providing auxiliary guidance and potentially reducing the risk of accidental injury.