Intraoperative applications of artificial intelligence for augmented parathyroid gland recognition: a narrative review.

Background and objective: Intraoperative parathyroid gland recognition is a key step during thyroidectomy to decrease the risk of postoperative hypocalcemia and during parathyroidectomy to distinguish normal and abnormal glands. Current methods for intraoperative identification rely largely upon visual identification. Recent investigation of methods such as near-infrared (NIR) autofluorescence and indocyanine green (ICG) for enhanced recognition have demonstrated steep learning curves. Artificial intelligence (AI) augmentation of all methods of parathyroid gland identification may improve intraoperative recognition rates and ultimately decrease rates of postoperative hypoparathyroidism. This narrative review aims to summarize the status of intraoperative application of AI for parathyroid gland recognition.

Methods: A systematic, comprehensive literature search was conducted using the search terms "artificial intelligence", "deep learning", "surgery", "parathyroid gland", and "parathyroid glands". Inclusion criteria included articles in English with the majority of the article devoted to intraoperative applications of AI on parathyroid gland recognition. Eleven studies were identified and included.

Key content and findings: Eight studies focused on utilizing AI intraoperatively to identify parathyroid glands from surrounding tissues. Three studies focused on using AI to predict abnormal from normal parathyroid glands. Five studies used NIR autofluorescence, two studies used visual recognition during open thyroidectomy, two studies used visual recognition during endoscopic thyroidectomy, one study used NIR autofluorescence with ICG angiography, and one study used coaxial dual-red-green-blue/near-infrared (dual-RGB/NIR) imaging system to identify parathyroid glands. Recall and precision scores for the models ranged from 50-95% and 72-94%, respectively. Four studies compared model performance with that of senior and junior surgeons and found that the models outperformed junior surgeons while performing comparably to senior surgeons.

Conclusions: AI augmentation of intraoperative parathyroid gland recognition demonstrates adequate accuracy results across a range of parathyroid gland recognition methods. Although these models are not currently available for widespread commercial use, the eventual integration into clinical practice may allow for enhanced intraoperative recognition of parathyroid glands, particularly in lower volume centers and for junior level surgeons.