AI dataset targets nerve injury risk during thyroid surgery 

A dataset of more than 18,000 annotated surgical images demonstrates the feasibility of artificial intelligence (AI)–based identification of the recurrent laryngeal nerve during endoscopic thyroidectomy, according to a study published in Scientific Data by lead authors Huaijin Zheng and Ruohan Cui, both of the Department of General Surgery, State Key Laboratory of Complex Severe and Rare Diseases, Peking Union Medical College Hospital, Chinese Academy of Medical Science and Peking Union Medical College in Beijing, China, and colleagues.

The recurrent laryngeal nerve (RLN) is vulnerable to injury in 3% to 8% of thyroidectomies and may result in complications such as vocal cord paralysis and aspiration pneumonia. 

The study describes ThyRLN-PUMCH, a curated dataset designed to support artificial intelligence (AI) models for intraoperative use. Researchers collected high-resolution surgical videos from 28 patients who underwent endoscopic thyroidectomy via chest-breast or trans-axillary approaches at a single tertiary center between 2020 and 2022. The RLN was confirmed intraoperatively in all cases, and procedures were performed by experienced surgeons using standardized equipment. 

Frames were extracted at four frames per second from video segments in which the RLN was visible. A total of 18,178 images were annotated by two endocrine surgeons and validated through a two-stage quality control process with independent expert verification. An additional 734 images without visible RLN were included for model testing. 

The dataset captured a range of intraoperative conditions, reflecting real-world surgical variability. Instrument occlusion was present in 47% of images and body fluid contamination in 3%, while 98% of images were acquired under bright lighting conditions and 2% were classified as low quality.  

To evaluate usability, investigators trained and tested two segmentation models: a convolutional neural network (DeepLabV3+) and a transformer-based model (Mask2Former). 

Using 9,998 images for training and 8,180 for testing, DeepLabV3+ achieved 64% recall, 65% precision, and a Dice score of 43%, while Mask2Former achieved 67% recall, 84% precision, and a Dice score of 47%. 

While the transformer-based model improved precision by nearly 20 percentage points compared with the baseline, recall gains were modest at 3 percentage points, highlighting highlighting persistent challenges in detecting small or partially obscured nerve structures. 

Limitations and Future Potential

The study does have several limitations. The dataset was derived from a single institution and included 28 cases, all performed using two endoscopic approaches, with no robotic or alternative techniques represented. Use of a single-imaging system may limit generalizability across operating environments. The study also did not quantify excluded frames during quality filtering, and model performance indicates room for improvement, particularly in recall.  

Still, this study gives clinicians a glimpse at the potential AI systems may hold in accurately identifying anatomical regions vulnerable to complications during thyroidectomies, making a strong case future research is warranted.

“The creation of this dataset carries profound implications for both surgical education and clinical practice [and] lays a solid foundation for the construction of future artificial intelligence-assisted surgical models,” the researchers wrote. “Furthermore, the dataset's structure facilitates research into temporal analysis of surgical workflows, instrument-tissue interaction modeling, and adaptive learning systems for anatomical variation handling.” 

The study was supported by National High Level Hospital Clinical Research Funding. The researchers reported no conflicts of interest. 

(Editor's Note: The researchers noted that the study manuscript will undergo further editing before final publication and that there may be errors present that affect the content.)

AACE Endocrine AI is published by Conexiant under a license arrangement with the American Association of Clinical Endocrinology, Inc. (AACE^®). The ideas and opinions expressed in AACE Endocrine AI do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.