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

Introduction

Artificial intelligence (AI) is defined as technology enabling computers or machines to perform tasks such as problem solving, object recognition, and decision making. Recent advances in AI have had wide ranging impacts on the medical field and all phases of surgical care. Preoperatively, AI assists in diagnostic tasks such as predicting benign and malignant skin lesions (1). Intraoperatively, models are being trained on visual stimuli to distinguish various pathologies, such as glioblastoma from meningiomas during neurosurgical procedures (2). AI has also been used to predict postoperative courses, such as the likelihood of surgical site infections (3).

Intraoperative recognition of parathyroid glands is a vital step of thyroidectomies. Parathyroid glands are similar in appearance to surrounding thyroid tissue, fat, and lymph nodes, posing difficulties to successful detection (4). The reported incidence of injury or removal resulting in postoperative hypoparathyroidism ranges from 19–38% for temporary and 0–3% for permanent hypoparathyroidism (5). Symptoms include neurologic, psychiatric, skeletal, neuromuscular, renal, and ocular manifestations with overall negative impacts on quality of life (4,5). As such, successful intraoperative recognition is necessary.

Advances in imaging modalities have allowed for preoperative localization of a majority of abnormal parathyroid glands in patients with hyperparathyroidism (6). A subset of patients remain, however, that either do not localize on preoperative imaging or have multi-gland disease in which sensitivity of imaging modalities decreases substantially (7). As such, intraoperative assessment of multiple parathyroid glands is often required. Differentiating normal from abnormal glands is paramount to the success of the procedure.

Current methods for intraoperatively identifying parathyroid glands include but are not limited to visual identification, near-infrared parathyroid autofluorescence (NIRAF), and indocyanine green (ICG).

Visual inspection is the predominant method of identification. Sensitivity ranges from 61–93%. Visual inspection is highly dependent upon experience level. Variation in size, color, and shape among parathyroid glands can make identification based on visual inspection alone difficult. Fine needle aspiration or frozen sections can be used as histopathologic confirmation of visual identification; however, this increases operative duration and can potentially damage normal tissue (8).

In NIRAF, either a probe or camera is used to stimulate fluorescence of the tissue in the operative field. Parathyroid glands emit a unique wavelength of fluorescence, allowing them to be distinguished from surrounding tissues (9). Two versions exist: an image-based system and a probe-based system. The image-based system, Fluorbeam-800, photographs the operative field and produces grayscale images developed using autofluorescence. Parathyroid glands are brighter compared to surrounding tissue. The probe-based system, PTeye, generates quantitative data including detection level and detection ratio after touching the tissue of interest. The standard threshold for parathyroid gland identification is a detection ratio greater than 1.2. Neither were developed utilizing AI or machine based learning technologies (10). While the probe based system provides quantitative values, the camera based system provides a fluorescent image requiring the surgeon to interpret the visualization intraoperatively (11). ICG is often given in combination with NIRAF and is used to examine perfusion to parathyroid glands (12). Limitations of all methods include the learning curve required for interpretation.

Augmenting intraoperative parathyroid gland recognition methods using AI may allow for efficient enhanced recognition of parathyroid glands. Multiple recent studies have developed AI models based on intraoperative images for identifying parathyroid glands. A recent review article by Apostolopoulos et al. summarized AI use for both preoperative and intraoperative parathyroid gland localization. Seven studies focusing on preoperative localization and six studies assessing intraoperative applications were included. Multiple AI models applied to both open and endoscopic procedures were included in the intraoperative studies. NIRAF was the main modality used for identification of parathyroid glands. Key limitations of AI use discussed here included lack of integration of clinical patient factors into the models, bias introduced by extracting images from a small patient population, and lack of a standardized method of validation, which will be discussed further below (13). Although the majority have yet to be truly tested intraoperatively, the results of the models suggest that intraoperative application would be feasible. Since its publication, multiple additional studies examining intraoperative AI uses for parathyroid gland localization have occurred. Herein, we discuss the status of intraoperative AI models for parathyroid gland recognition. We present this article in accordance with the Narrative Review reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-165/rc).

Methods

A systematic comprehensive literature search was performed to identify articles pertaining to intraoperative use of AI to augment parathyroid gland recognition. A PubMed/MEDLINE search between 2000–2025 was conducted using a combination of the following search terms: “artificial intelligence”, “deep learning”, “surgery”, “parathyroid gland”, and “parathyroid glands”. The inclusion criteria were articles published in English that explored the intraoperative application of AI pertaining to the identification of parathyroid glands. Articles were excluded if the majority of the focus of the article was either not related to intraoperative applications of AI or not related to parathyroid glands (Table 1). Eleven articles were included and their evaluations of the role of intraoperative AI on parathyroid gland recognition are summarized in Table 2.

Machine learning and deep learning models

Machine learning is a form of AI that utilizes pattern recognition to enable the program to learn content and ultimately make predictions and recommendations (24). Supervised learning involves training the model on a category of labelled data points such that it can then make predictions when presented with new datasets of similar content. Unsupervised learning uses non labeled data points to create an algorithm that groups similar characteristics of new datasets together without a set correct answer (25). The majority of the articles discussed below utilized supervised learning to create a program that predicts the location of parathyroid glands.

Deep learning is a subset of machine learning that uses neural networks to learn complex patterns (26). Deep neural networks have multiple layers integrated into the model allowing them to understand more complex topics (27). Examples of deep neural networks can be seen in a range of roles such as breast screening models to interpretation of endomyocardial biopsy results (26,27). A portion of the articles discussed below utilized deep learning and deep neural networks for analysis of parathyroid glands.

Accuracy of the AI models below is largely reported with precision and recall scores. Precision is a measure of correct predictions and is calculated with the formula (true positives)/(true positives + false positives). Recall is a measure of correct identification of relevant data points and is calculated with the formula (number of positive samples correctly classified)/(total number of positive samples). Precision is similar to positive predictive value while recall is similar to sensitivity (28).

Applications of intraoperative AI for parathyroid gland recognition

Impacts of AI on parathyroid gland recognition during endoscopic thyroidectomy

The prevalence of minimally invasive video assisted thyroidectomy has risen over recent years, however a substantial number of cases are required to become proficient in this approach (29). Postoperative hypoparathyroidism reports range from 1–22% for transient hypoparathyroidism and 1–2% for permanent hypothyroidism after endoscopic thyroidectomy (30). Wang et al. [2022] created a deep learning model called Artificial Intelligence Model for Parathyroid Gland Recognition (PTAIR) using video clips from endoscopic surgery through the bilateral areola approach in an attempt to facilitate parathyroid gland recognition during the learning curve of this approach. The model was trained with an internal learning cohort consisting of 1,700 images from 166 endoscopic thyroidectomy videos and tested on an external validation cohort consisting of 20 full length videos from a senior surgeon. Parathyroid glands were identified through a series of video review conducted sequentially by two junior surgeons and then two senior surgeons. Of note, one of the senior surgeons who reviewed the images and was part of the testing cohort provided the 20 full length videos used in the external cohort group. PTAIR was subsequently compared with both junior and senior surgeon recognition of parathyroid glands. PTAIR demonstrated 88.7% precision and 92.3% recall. PTAIR recognized 96.9% of parathyroid glands while junior surgeons recognized 71.9% and senior surgeons recognized 87.5%. PTAIR performed significantly better than junior surgeons (14).

The same authors subsequently modified PTAIR to create and test PTAIR 2.0 in Wang et al. [2024]. Here, the deep learning model was modified to predict parathyroid gland location and provide ischemia alerts in addition to recognizing parathyroid gland location. This model again was trained on an internal validation cohort using 413 video clips from endoscopic surgery in 838 patients, tested on an external validation cohort consisting of 54 patients, and compared with junior and senior surgeon performance. Parathyroid glands were identified in the same manner as PTAIR. Precision scores for parathyroid gland identification, early prediction, and ischemia alerts were 98.9%, 94.1%, and 92.1% in the internal validation cohort. In the external validation cohort, PTAIR 2.0 had a significantly higher prediction rate, identification rate, and ischemic assessment rate compared with junior surgeons (63.64% vs. 11.36%, 94.51% vs. 64.84%, 84% vs. 52%, respectively). PTAIR 2.0 also had a significantly higher prediction rate compared with senior surgeons (63.64% vs. 34.09%). PTAIR 2.0 was thus concluded to have a higher overall performance than junior surgeons, as well as a higher rate of prediction compared with senior surgeons (15).

A limitation of these two studies is that this model is only applicable for an endoscopic approach. The results cannot be generalized to open approaches. Additionally, in both studies parathyroid gland presence was confirmed by senior surgeon visual validation rather than histopathology examination, leaving room for error as discussed previously. The authors acknowledged and accounted for this by reporting a 90% confidence level. A major limitation is that all work with this system has been conducted by one single center research group. This raises multiple concerns, including the risk of confirmation bias, selection bias, and generalizability. PTAIR was trained on a relatively limited number of patients. Although the authors extracted a substantial number of video clips from that sample, the lower number of patients exposes the model to a limited amount of parathyroid variability. Although the patient sample size was larger in PTAIR 2.0, the number of images utilized to train the model was substantially lower, echoing the same problem. This also introduces the risk of overfitting with machine learning, which occurs when the model becomes highly attuned to the training set rather than the general patterns and can lead to difficulty identifying targets in new data. Ultimately, the PTAIR and PTAIR 2.0 models did perform comparably to senior level surgeons, suggesting an acceptable rate of performance that could allow for intraoperative application. Intraoperative application would be feasible if the model can be integrated with endoscopic systems to allow for results in real time. Work with these models by independent multi-institutional groups would enhance external validity and increase the sample size variability that the models are trained and tested on, mitigating the overfitting effect. This should occur prior to integrating the technology into endoscopic systems.

Augmentation of parathyroid gland recognition with NIRAF with AI

Parathyroid glands possess intrinsic fluorescence properties that when stimulated give off a specific wavelength, aiding in their identification. Previous studies have utilized NIRAF to identify parathyroid glands intraoperatively, with parathyroid glands demonstrating higher fluorescence intensity than surrounding tissues (31). There have been reports of increased early parathyroid gland identification, as well as lower rates of temporary postoperative hypocalcemia with NIRAF (32). A limitation of NIRAF, particularly when using image-based models, is the learning curve required to interpret the autofluorescence based identification. While the probe-based modalities provide quantitative data, the image-based modalities rely upon surgeon ability to distinguish brighter tissues from the surrounding tissues in the photographs. Although this is used in combination with visual identification, it is subject to many of the same pitfalls. As such, the addition of AI identification and prediction would further facilitate the intraoperative use of image-based NIRAF (21,33,34).

Avci et al. [2022] examined AI interpretation of intraoperative NIRAF images of patients undergoing open parathyroidectomy and thyroidectomy. Deep learning was used to train an AI platform to recognize parathyroid autofluorescence signals. A total of 466 images from 197 patients were utilized, and 80% were used for training, 10% for testing, and 10% for validation. Recall was 90.5% and precision was 95.7%. The model correctly identified 91.9% of parathyroid glands, with identification of the glands confirmed by senior surgeon visual interpretation. False positives largely correlated to brown fat, thyroid cysts, and surgical gloves. False negatives correlated with darker and deeper parathyroid glands. Of note, the time from image generation to model prediction was approximately 1 minute (16).

Like the above studies, a limitation of this study is the lack of histopathologic confirmation of parathyroid gland location. Additionally, images were extracted from a single institution patient population limiting the variability the model was exposed to. This model did, however, demonstrate adequate precision and recall scores with short processing time, suggesting it as a feasible intraoperative modality to assist in parathyroid gland recognition without significantly increasing operative times. Much like visual inspection, the model was subject to false positives and negatives. The authors suggested that utilizing this technology at multiple points of the dissection may help combat the false positive and negative rates. Although the processing time was short, the need for repeated use during the procedure runs the risk of prolonging operative time. Lastly, there is no commercially available model integrating NIRAF with this AI platform.

False positives with NIRAF can limit its intraoperative assistance. Yu et al. [2024] created a machine learning model for recognition of parathyroid glands, compared the model performance with senior and junior surgeons, and analyzed false positive rates among all groups. The model was trained on intraoperative NIRAF images where parathyroid glands were labelled by a professional surgeon. A total of 452 images from 72 patients were divided into internal and external validation cohorts in a ratio of 8:2. Precision and recall for the internal validation cohort were 83.5% and 57.8%. Precision and recall for the external validation cohort were 74.5% and 50.5%. Total recognition of parathyroid glands was comparable between the model and senior surgeons (85.2% vs. 83.5%) while the model performed significantly better than junior surgeons (85.2% vs. 70.4%). False positive rates in the internal validation cohort for the model, senior surgeons, and junior surgeons were 12.5%, 9.1%, and 12.5%. False positive rates in the external validation cohort for the model, senior surgeon, and junior surgeon were 11.3%, 4.2%, and 2.8%. The most common tissue falsely identified as a parathyroid gland was deep adipose tissue both for the model and the surgeons (8).

Similar to PTAIR and PTAIR 2.0, the model had higher recognition of parathyroid glands than junior level surgeons and demonstrated comparable performance to senior level surgeons. However, false positive rates were higher in the model than in either surgeon group. This study utilized visual identification in combination with intraoperative parathyroid hormone levels to confirm parathyroid gland location but would be strengthened by histopathologic confirmation. Additionally, this study also required extraction of multiple images from a small patient population, limiting the variability the model was exposed to. The decline in precision and recall scores in the external validation cohort suggests that overfitting may have occurred with this model such that it did not perform as well on the independent data set. Finally, this platform is also not integrated with a NIRAF camera in a commercially available manner.

AI integration with NIRAF enhances outcomes compared to NIRAF alone largely for junior surgeons. The benefits of utilization by a senior surgeon are marginal at best in the above studies. Intraoperative application for both models would require the presence of a NIRAF camera, the respective computer system, and a display monitor. The AI models would need to be integrated with NIRAF technology and require regulatory approval as well as institutional buy-in for widespread use. Multi-center trials involving larger patient populations to expose the models to increased variability would be beneficial next steps prior to integration with commercially available NIRAF technology. These models can be applied to open surgery, as this allows for sufficient exposure of the surgical field to ensure adequate NIRAF images are captured, however they are less applicable in minimally invasive approaches.

AI and ICG detection of parathyroid glands

While NIRAF can distinguish parathyroid glands from other tissues, it does not always correlate with adequate perfusion and function. Combining NIRAF with ICG allows for the evaluation of both location and perfusion of the parathyroid glands. McEntee et al. [2024] trained a simple regression model using near-infrared (NIR) ICG parathyroid gland angiography videos. The model was trained on 37 videos and tested on 22 videos including four intraoperatively in live time. Precision was 90% and recall was 93.3%. Additionally, results were generated within 5 minutes of image capture, making it a feasible efficient method for intraoperative utilization (17).

This study was the only study to report intraoperative testing of the proposed AI model and demonstrated only a minimal addition to operative time. However, the sample size was small for both training the model, limiting the variability it was exposed to, and testing the model, limiting the evaluation of its performance on a new dataset. Additionally, only a small portion was tested intraoperatively. While it did demonstrate adequate accuracy scores, the results should be interpreted with caution given the small sample size. This model would also benefit from further testing with a larger, multi-institutional patient population.

AI and co-axial excitation dual-red-green-blue/near-infrared (dual-RGB/NIR) paired imaging system for recognizing parathyroid glands

As discussed above, there are limitations to NIRAF imaging that impact its accuracy. Kim et al. [2022] created a clinical imaging prototype consisting of a combination of a coaxial, collimated fiber laser module with a co-registered dual-RGB/NIR camera model to evaluate if combining RGB and NIR imaging would enhance its success. They first conducted benchtop testing to optimize the settings of the camera and then subsequently utilized it intraoperatively. The dataset was composed of 1,287 images from six patients. Four patients’ images were used to train the deep learning model using You Only Look Once Version 5 (YOLOv5) and Faster Region-based Convolutional Neural Network (Faster R CNN) and two patients’ images were used to evaluate model performance. Both thyroidectomies and parathyroidectomies were included. Excised parathyroid glands were sent for histopathologic confirmation while visual inspection was used to confirm non-excised glands. The mean average precision was 94.7% for YOLOv5 and 89.8% for Faster R CNN for combination imaging with dual-RGB/NIR. Both models were tested with NIR imaging alone and RGB alone and performed better with combined dual-RGB/NIR (18).

Although the models demonstrated adequate mean average precision scores, the model was trained on a very small sample of patients. Although many images were extracted, the variability that exists within parathyroid glands may not be captured with the limited number of patients. This model is also at risk for overfitting. Additionally, the specific imaging model used by these authors is not commercially available and would need to be tested on a larger scale with reproducible results prior to widespread use intraoperatively.

AI use for detection of abnormal parathyroid glands

Prior studies have demonstrated that normal and abnormal parathyroid gland NIRAF signals demonstrate differences in intensity and homogeneity (21,33-35). It has also been demonstrated that NIRAF may identify approximately 20% of parathyroid glands prior to the surgeon, which would facilitate early recognition (32).

Akbulut et al. [2021] aimed to develop an AI model to distinguish normal from abnormal parathyroid glands. Quantitative values for autofluorescence intensity, heterogeneity index, and parathyroid gland volume were assigned using a third party software to intraoperative NIRAF images obtained by experienced endocrine surgeons during open parathyroidectomy. The authors did not report the patient or image sample size. This data was used to develop models and deep neural network decision trees trained with machine learning to predict normal from abnormal glands and distinguish the type of gland abnormality. Decision trees were created either utilizing autofluorescence intensity and heterogeneity index or autofluorescence intensity, heterogeneity index, and parathyroid gland volume. Accuracy of the two-parameter model was 88.5% and accuracy of the three-parameter model was 97.4%. The accuracy of the models to predict the type of abnormal parathyroid gland were 69% for the two-parameter model and 84% for the three-parameter model. Precision scores using the three-parameter model were highest for normal glands and single adenomas and lower for double adenomas and hyperplasia (97.3%, 93.4%, 41.9%, and 38.1%, respectively) (20). The authors are currently conducting an additional larger study to evaluate the success of the model when utilized by surgeons of all experience levels.

Abnormal parathyroid glands were sent for histopathologic confirmation, and intraoperative parathyroid hormone was monitored according to the Miami criteria, which provides additional support for correct identification of those glands. Normal appearing glands were again confirmed by visual inspection. A major limitation of this study is that the authors did not explicitly list their sample size, making it difficult to interpret the variability the model was trained on. Additionally, intraoperative application of this model would require many physical components, including the NIRAF system, the third-party software used to assign quantitative values to the photos, and integration of the AI model. Given the substantial resources needed to utilize this model, it may present more barriers to intraoperative use. Larger scale trials with multiple institutions would be required before widespread use of this model.

Avci et al. [2022] examined the ability of AI to differentiate abnormal from normal parathyroid glands. In this study, 906 intraoperative NIRAF images from 303 patients were obtained from four gland exploration parathyroidectomies for primary hyperparathyroidism, and 80% of images were used for training, 10% for testing, and 10% for validation. Both recall and precision were 89% for differentiating normal and abnormal glands. Precision for prediction of an abnormal gland was 91% and precision for prediction of a normal gland was 88%. They concluded that AI interpretation of NIRAF images would be beneficial during four gland exploration for patients who may not have localized preoperatively or in whom intraoperative parathyroid hormone does not drop appropriately (19).

Histopathologic confirmation with intraoperative parathyroid hormone monitoring was used to confirm abnormal glands, strengthening the identification of abnormal parathyroid tissue. The non-excised glands were assumed to be normal parathyroid glands based on visual inspection alone. Although the sample size here was larger compared to other reported studies, this remained a single institution study, limiting its generalizability and variability for training the model. Again, this model would benefit from multi-institutional investigation.

Akgun et al. [2025] created a deep learning model to assist in the intraoperative recognition of parathyroid glands and the differentiation of normal from abnormal glands. This same group had previously created a visual deep learning model to differentiate parathyroid tissue from surrounding tissue using ex vivo NIRAF images, demonstrating the ability of AI models to predict parathyroid gland presence based on NIRAF (36). They subsequently applied the same approach to in vivo application. A prospective single center study was conducted including patients with primary hyperparathyroidism undergoing parathyroidectomy or normocalcemic patients undergoing total thyroidectomy or lobectomy with intraoperative NIRAF image capture. In vivo NIRAF images were used to create a neural network to predict the location of parathyroid glands as well as normal vs. abnormal parathyroid glands. In vivo images of abnormal glands during parathyroidectomy were confirmed using intraoperative parathyroid hormone values and histopathology results. A total of 1,506 normal and 597 abnormal parathyroid glands were analyzed with images. Both precision and recall of the model were 83.3%. The authors proposed that eventual widespread use of this technology could aid in assisting lower volume centers with accurate parathyroid gland detection intraoperatively (21).

This study utilized a large number of parathyroid gland images, both normal and abnormal, although the training and testing breakdown was not clearly described. The external validity of this model remains limited given that the AI model was trained on images only from a single institution. Only the parathyroid glands identified in NIRAF photos from parathyroidectomies were confirmed on pathologic diagnosis. The remainder were confirmed with visual identification. While the authors predicted that this could assist in lower volume centers, this technology would first require multi-institutional testing and integration with NIRAF systems. The necessary financial and institutional investments in AI integrated NIRAF technology may limit the ultimate intraoperative use.

All three of the above studies confirmed abnormal parathyroid glands with histopathologic diagnosis and intraoperative parathyroid hormone monitoring, strengthening the evidence that the models were trained on true abnormal glands. An important limitation of all NIRAF studies is that no current large database of intraoperative NIRAF images exists as a comparison point for labelling images. As such, it continues to rely upon surgeon and research group experience level, likely accounting for some of the variability in the above model accuracy scores. Additionally, these studies were all conducted as single center projects at the Cleveland Clinic, further limiting the sample size and variability the models were trained on. This ultimately limits the external validity and generalizability of this technology. Finally, no current NIRAF systems are integrated with an AI model. These models would benefit from multi center trials to demonstrate adequate precision and recall, as well as intraoperative practicability prior to investing in integrating these technologies and seeking regulatory approval.

Augmentation of visual identification of parathyroid glands during open thyroidectomy

Different approaches to thyroidectomy pose different risks to parathyroid gland preservation. Lee et al. [2024] aimed to create an AI model that would be generalizable to different operative approaches. Video clips were obtained during open thyroidectomy by a single surgeon utilizing both commercial and laparoscopic cameras and used to create multiple AI models with different types of data augmentation. Parathyroid glands were confirmed using fine needle aspiration. A total of 152 images extracted from 150 patients were used. The analysis ultimately included an object identification model, a model utilizing geometric transformation aimed at precise localization and identification of the parathyroid glands, and one using image inpainting aimed at creating new images reconstructing the missing regions of the intraoperative images after tissue removal. Average precision scores were overall similar in each model (77% object identification, 79% geometric transformation, 78.6% image inpainting) when used on the training set of open images. In the external validation cohort, average precision scores were highest for the image inpainting model (33% object identification, 37.7% geometric transformation, 45.9% image inpainting).

A major limitation of this study was all intraoperative data came from a single surgeon, limiting its generalizability. The authors are planning for a multi-institutional randomized control trial to address this and assess model performance against surgeon performance (22). Additionally, the image sample size is overall small, further contributing to limited generalizability. Overfitting likely occurred given the decline in external validity scores. The use of both open and minimally invasive approaches would enhance the applicability of this model and potentially incentivize its use over models limited to one operative modality.

Intraoperative perspectives are dynamic and changing constantly from one moment to the next. The above studies focused largely on static viewpoints obtained across multiple patients and procedures. Sang et al. [2024] utilized intraoperative videos to train a deep learning model for predicting the location of parathyroid glands in 65 patients undergoing open thyroidectomy. Parathyroid glands were labelled by a senior surgeon with more than 20 years of experience. The model was trained on 110 videos with 9,944 frames in the internal validity group and tested on an external cohort consisting of 15 videos with 1,500 frames. The model was compared with senior and junior surgeons. Precision and recall were 84.9% and 81.3% for the internal validation cohort. When tested on the external validation cohort, the model achieved precision and recall scores of 89.5% and 70.8%. In the external validation cohort, two junior surgeons achieved precision scores of 75% and 80% and recall scores of 81.8% and 72.7%. Two senior surgeons demonstrated precision scores of 70% and 84.6% and recall scores of 63.6% and 84.6%. Precision scores were thought to be higher for the model as it was providing accurate predictions, as compared with surgeons who tended to propose more possibilities. Additionally, since video clips were used, the authors were able to demonstrate that the AI model could track the parathyroid glands throughout the procedure (23).

Patient videos used to train the model all originated from a single institution. As such, the variability of different glands and patient populations may not be adequately captured. Additionally, the practical application of intraoperative videos during open procedures is overall limited and would likely add to operative time.

Conclusions

Visual recognition remains the main method for identification of parathyroid glands, both during thyroidectomy to prevent postoperative hypoparathyroidism and during parathyroidectomy to detect abnormal glands. Multiple methods remain under investigation to augment the identification of parathyroid glands, including NIR autofluorescence signals and ICG. The above studies demonstrated overall adequate performance by AI models to detect parathyroid glands and distinguish normal from abnormal glands based on intraoperative images. Multiple studies demonstrated enhanced performance by the AI models compared to junior level surgeons with comparable performance to senior level surgeons. While not yet commercially available on a large scale, the eventual implementation of such technology would assist in transferring senior level surgeon experience at high volume centers to junior level surgeons and lower volume centers.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-165/rc

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-165/prf

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-165/coif). K.N.P. serves as an unpaid editorial board member of Gland Surgery from December 2024 to November 2026. The other author has no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017;542:115-8. [Crossref] [PubMed]
2. Gupta A, Singla T, Chennatt JJ, et al. Artificial intelligence: A new tool in surgeon's hand. J Educ Health Promot 2022;11:93. [Crossref] [PubMed]
3. Soguero-Ruiz C, Fei WM, Jenssen R, et al. Data-driven Temporal Prediction of Surgical Site Infection. AMIA Annu Symp Proc 2015;2015:1164-73.
4. Pasieka JL, Wentworth K, Yeo CT, et al. Etiology and Pathophysiology of Hypoparathyroidism: A Narrative Review. J Bone Miner Res 2022;37:2586-601. [Crossref] [PubMed]
5. Orloff LA, Wiseman SM, Bernet VJ, et al. American Thyroid Association Statement on Postoperative Hypoparathyroidism: Diagnosis, Prevention, and Management in Adults. Thyroid 2018;28:830-41. [Crossref] [PubMed]
6. Uludag M. Preoperative Localization Studies in Primary Hyperparathyroidism. Sisli Etfal Hastan Tıp Bul 2019;53:7-15. [Crossref] [PubMed]
7. Siperstein A, Berber E, Barbosa GF, et al. Predicting the success of limited exploration for primary hyperparathyroidism using ultrasound, sestamibi, and intraoperative parathyroid hormone: analysis of 1158 cases. Ann Surg 2008;248:420-8. [Crossref] [PubMed]
8. Yu F, Sang T, Kang J, et al. An automatic parathyroid recognition and segmentation model based on deep learning of near-infrared autofluorescence imaging. Cancer Med 2024;13:e7065. [Crossref] [PubMed]
9. Kiernan CM, Solόrzano CC. New Techniques for Intraoperative Parathyroid Localization. Adv Surg 2023;57:209-23. [Crossref] [PubMed]
10. Solórzano CC, Thomas G, Berber E, et al. Current state of intraoperative use of near infrared fluorescence for parathyroid identification and preservation. Surgery 2021;169:868-78. [Crossref] [PubMed]
11. Thomas G, Squires MH, Metcalf T, et al. Imaging or Fiber Probe-Based Approach? Assessing Different Methods to Detect Near Infrared Autofluorescence for Intraoperative Parathyroid Identification. J Am Coll Surg 2019;229:596-608.e3. [Crossref] [PubMed]
12. van den Bos J, van Kooten L, Engelen SME, et al. Feasibility of indocyanine green fluorescence imaging for intraoperative identification of parathyroid glands during thyroid surgery. Head Neck 2019;41:340-8. [Crossref] [PubMed]
13. Apostolopoulos ID, Papandrianos NI, Papageorgiou EI, et al. Artificial Intelligence Methods for Identifying and Localizing Abnormal Parathyroid Glands: A Review Study. Mach Learn Knowl Extr 2022;4:814-26.
14. Wang B, Zheng J, Yu JF, et al. Development of Artificial Intelligence for Parathyroid Recognition During Endoscopic Thyroid Surgery. Laryngoscope 2022;132:2516-23. [Crossref] [PubMed]
15. Wang B, Yu JF, Lin SY, et al. Intraoperative AI-assisted early prediction of parathyroid and ischemia alert in endoscopic thyroid surgery. Head Neck 2024;46:1975-87. [Crossref] [PubMed]
16. Avci SN, Isiktas G, Berber E. A Visual Deep Learning Model to Localize Parathyroid-Specific Autofluorescence on Near-Infrared Imaging : Localization of Parathyroid Autofluorescence with Deep Learning. Ann Surg Oncol 2022; Epub ahead of print. [Crossref]
17. McEntee PD, Greevy JE, Triponez F, et al. Parathyroid gland identification and angiography classification using simple machine learning methods. BJS Open 2024;8:zrae122. [Crossref] [PubMed]
18. Kim Y, Lee HC, Kim J, et al. A coaxial excitation, dual-red-green-blue/near-infrared paired imaging system toward computer-aided detection of parathyroid glands in situ and ex vivo. J Biophotonics 2022;15:e202200008. [Crossref] [PubMed]
19. Avci SN, Isiktas G, Ergun O, et al. A visual deep learning model to predict abnormal versus normal parathyroid glands using intraoperative autofluorescence signals. J Surg Oncol 2022;126:263-7. [Crossref] [PubMed]
20. Akbulut S, Erten O, Kim YS, et al. Development of an algorithm for intraoperative autofluorescence assessment of parathyroid glands in primary hyperparathyroidism using artificial intelligence. Surgery 2021;170:454-61. [Crossref] [PubMed]
21. Akgun E, Ibrahimli A, Berber E. Near-Infrared Autofluorescence Signature: A New Parameter for Intraoperative Assessment of Parathyroid Glands in Primary Hyperparathyroidism. J Am Coll Surg 2025;240:84-93. [Crossref] [PubMed]
22. Lee JH, Ku E, Chung YS, et al. Intraoperative detection of parathyroid glands using artificial intelligence: optimizing medical image training with data augmentation methods. Surg Endosc 2024;38:5732-45. [Crossref] [PubMed]
23. Sang T, Yu F, Zhao J, et al. A novel deep learning method to segment parathyroid glands on intraoperative videos of thyroid surgery. Front Surg 2024;11:1370017. [Crossref] [PubMed]
24. Hashimoto DA, Rosman G, Rus D, et al. Artificial Intelligence in Surgery: Promises and Perils. Ann Surg 2018;268:70-6. [Crossref] [PubMed]
25. Kufel J, Bargieł-Łączek K, Kocot S, et al. What Is Machine Learning, Artificial Neural Networks and Deep Learning?-Examples of Practical Applications in Medicine. Diagnostics (Basel) 2023;13:2582. [Crossref] [PubMed]
26. Tsai KJ, Chou MC, Li HM, et al. A High-Performance Deep Neural Network Model for BI-RADS Classification of Screening Mammography. Sensors (Basel) 2022;22:1160. [Crossref] [PubMed]
27. Jiao Y, Yuan J, Sodimu OM, et al. Deep Neural Network-Aided Histopathological Analysis of Myocardial Injury. Front Cardiovasc Med 2021;8:724183. [Crossref] [PubMed]
28. Andersen ES, Birk-Korch JB, Hansen RS, et al. Monitoring performance of clinical artificial intelligence in health care: a scoping review. JBI Evid Synth 2024;22:2423-46. [Crossref] [PubMed]
29. Del Rio P, Sommaruga L, Cataldo S, et al. Minimally invasive video-assisted thyroidectomy: the learning curve. Eur Surg Res 2008;41:33-6. [Crossref] [PubMed]
30. Akritidou E, Douridas G, Spartalis E, et al. Complications of Trans-oral Endoscopic Thyroidectomy Vestibular Approach: A Systematic Review. In Vivo 2022;36:1-12. [Crossref] [PubMed]
31. Falco J, Dip F, Quadri P, et al. Cutting Edge in Thyroid Surgery: Autofluorescence of Parathyroid Glands. J Am Coll Surg 2016;223:374-80. [Crossref] [PubMed]
32. Kose E, Rudin AV, Kahramangil B, et al. Autofluorescence imaging of parathyroid glands: An assessment of potential indications. Surgery 2020;167:173-9. [Crossref] [PubMed]
33. Benmiloud F, Rebaudet S, Varoquaux A, et al. Impact of autofluorescence-based identification of parathyroids during total thyroidectomy on postoperative hypocalcemia: a before and after controlled study. Surgery 2018;163:23-30. [Crossref] [PubMed]
34. Benmiloud F, Godiris-Petit G, Gras R, et al. Association of Autofluorescence-Based Detection of the Parathyroid Glands During Total Thyroidectomy With Postoperative Hypocalcemia Risk: Results of the PARAFLUO Multicenter Randomized Clinical Trial. JAMA Surg 2020;155:106-12. [Crossref] [PubMed]
35. Kose E, Kahramangil B, Aydin H, et al. Heterogeneous and low-intensity parathyroid autofluorescence: Patterns suggesting hyperfunction at parathyroid exploration. Surgery 2019;165:431-7. [Crossref] [PubMed]
36. Akgun E, Uysal M, Avci SN, et al. The use of artificial intelligence to detect parathyroid tissue on ex vivo specimens during thyroidectomy and parathyroidectomy procedures using near-infrared autofluorescence signals. Surgery 2024;176:1396-401. [Crossref] [PubMed]