Computer Vision-Based Recognition of Parathyroid Glands and Recurrent Laryngeal Nerves in Endoscopic Surgery

Identification of the parathyroid glands and recurrent laryngeal nerves is an important component of safe thyroid surgery. Parathyroid glands are small and may resemble surrounding fat, lymph nodes, or soft tissue. The recurrent laryngeal nerve has a complex anatomical course and may be partially covered by fascia or insufficiently exposed in some cases, which may increase the difficulty of intraoperative recognition. Computer vision-based artificial intelligence models may help recognize these structures in endoscopic surgical images, but their performance in multicenter real-world surgical settings requires further validation. Previous PTAIR studies mainly focused on parathyroid gland recognition in endoscopic thyroid surgery videos. Based on these prior studies, the present study will further evaluate the PTAIR model in a prospective multicenter setting, with a focus on recognition of parathyroid glands and recurrent laryngeal nerves. This stage of the study is designed as a model performance validation study and will not evaluate whether PTAIR changes intraoperative surgical decisions, parathyroid management, or postoperative clinical outcomes. During endoscopic thyroid surgery, PTAIR, junior physicians, and senior physicians will identify parathyroid glands and recurrent laryngeal nerves based on the same intraoperative endoscopic visual field. All time-related outcomes will use the video recorder timeline as the unified reference. The study will record the first recognition time of the recurrent laryngeal nerve and parathyroid gland by PTAIR, junior physicians, and senior physicians, as well as the continuous recognition time of these structures by PTAIR. Expert-reviewed annotations will be used as the reference for subsequent model performance evaluation and confirmation of recognition results. The study will also evaluate PTAIR performance under different exposure conditions, including cases with fascial coverage, incomplete exposure, or complex surgical field conditions. Multicenter video data will be used to assess model performance across different centers, surgeons, equipment settings, and operative field conditions, and may provide data for future model improvement and iteration.