The current landscape of artificial intelligence in the diagnosis of thyroid nodules: a surgeon’s perspective

It is sometimes difficult to fathom that the invention of penicillin by Sir Alexander Fleming occurred almost a century ago and represented one of the most influential medical breakthroughs of the 1920’s. Since that time, the many advances in medicine that have improved our understanding of disease pathology and patient care are truly remarkable and reflect the medical community’s enduring commitment to improving patient outcomes. Currently, the field of medicine is entering a new phase of the information age, one that is shaping the role and application of artificial intelligence (AI) in patient care. Potential uses of AI are broad, ranging from facilitating disease diagnosis to pilot applications in robotic surgical technology (1). The development and optimal clinical application of AI remain works in progress and have prompted ongoing study within the medical community.

In their study, Edström and colleagues aimed to evaluate the accuracy of a recently developed AI-based diagnostic ultrasound tool designed to stratify the risk of thyroid nodule malignancy independent of user experience (2). “S-Detect for Thyroid”, a deep learning-based AI program, has previously demonstrated encouraging results in assisting clinicians with diagnosing and risk-stratifying thyroid nodules—particularly amongst less experienced users (3-5). The authors designed a clinical trial involving twenty users who evaluated the same five patients with thyroid nodules to determine whether AI technology improved nodule assessment and recommendations for fine-needle aspiration (FNA) biopsy. The study had two primary outcomes: (I) accuracy in stratifying malignancy risk (“S-Detect diagnostic accuracy”) and (II) biopsy recommendation before and after AI assistance. Outcomes were compared among eight medical students, three junior ultrasound registrars, and nine senior (experienced) ultrasound registrars.

Overall, the results of the study were underwhelming across all experience levels. For the first outcome, diagnostic accuracy with AI assistance was observed in 71% of participants and did not differ significantly amongst groups (students: 74%, ultrasound novices: 60%, ultrasound experienced: 72%; P=0.30). The technology accurately identified benign pathology, including cysts and follicular adenomas, regardless of user experience in 100% of cases. However, performance was substantially poorer for goiter (55%) and papillary thyroid carcinoma (15%). Regarding the second outcome, the AI-based S-Detect system did not improve users’ ability to determine whether a nodule required biopsy, with overall recommendation accuracy remaining unchanged at 69% before and after AI implementation. Although accuracy improved among ultrasound novice users (60% to 72%), it declined among medical students (74% to 67%). Notably, changes in biopsy recommendations resulted in four true-positive nodules being reclassified as false negatives and three true-negative nodules being reclassified as false positives.

AI assistance in stratifying thyroid nodules, independent of user experience, has potential clinical value. However, the results of this study indicate that further development is required before widespread clinical adoption. The authors appropriately conducted a pilot study to assess whether AI could enhance diagnostic performance among trainees. While the technology did not significantly improve diagnostic accuracy, it also did not consistently worsen performance, highlighting the current level of algorithmic sophistication. These findings suggest that the algorithm is capable of recognizing important morphological features and may approach clinical utility with further refinement and improved pattern recognition. The study also introduces several important concepts related to AI technology that warrant discussion.

Although this study utilized a commercially available AI-based deep learning platform, the use of AI in thyroid nodule diagnosis is not novel. Multiple systems have been developed to stratify malignancy risk using the Thyroid Imaging Reporting and Data System (TIRADS) framework to guide biopsy decisions. Several studies have demonstrated that AI-based tools perform comparably to, or even improve upon, diagnoses made by experienced radiologists. By improving nodule specificity, AI-assisted diagnosis has been reported to reduce unnecessary FNA biopsies by approximately 48–60% (6,7). These encouraging findings have been incorporated into societal guidelines, particularly for identifying benign nodules in low-risk populations.

Despite advances in risk stratification, AI-based characterization of other aspects of thyroid ultrasound remains limited. While many systems have been trained to identify papillary thyroid carcinoma, their ability to detect other malignancies (such as follicular or medullary carcinoma) is less robust. Additionally, the characterization of noninvasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) and the follicular variant of papillary thyroid carcinoma (FVPTC) remains under investigation. Another evolving area is the application of AI to cytopathologic and histopathologic classification—including Bethesda categorization—which remains in development and highly user dependent. Furthermore, AI-based detection of lymph node metastases is challenging due to subtle imaging features that require further refinement. Although AI technology has demonstrated impressive progress, its application to the comprehensive diagnosis and work-up of thyroid cancer remains a work in evolution.

Concerns regarding reliability also limit the current use of AI in thyroid pathology. One of the primary criticisms of medical AI is the lack of transparency inherent to “black box” models, in which clinicians are unable to fully understand the algorithmic reasoning behind generated recommendations (8). This issue is particularly relevant in thyroid disease where management often involves observation or de-escalation and decisions regarding surgical extent, radioactive iodine use, and surveillance require nuanced clinical judgment. Additionally, many AI models lack prospective, multicenter validation. As demonstrated in this study, although S-Detect had previously shown diagnostic utility, its performance was not generalizable across varying patient pathologies and user experience levels, limiting its clinical applicability. While continued technological advancements may address these issues, they remain areas of active debate.

Accountability represents another unresolved limitation of AI implementation. It remains unclear where responsibility lies when clinical decisions influenced by AI result in error—whether with the algorithm developer or the clinician. At present, medical decision-making remains the responsibility of the provider, and as demonstrated in this study, AI assistance did not substantially alter user decisions. However, as AI becomes more autonomous, questions surrounding liability will become increasingly important. Future clinical integration may necessitate formal consent processes or disclosures, though this remains an unresolved ethical and legal challenge.

Ethical and equity concerns further complicate AI adoption. Current AI models are highly dependent on the datasets used for training, which may underrepresent certain patient populations and introduce bias. This raises concerns regarding health disparities, particularly when algorithms are developed within limited institutional or regional datasets that may not generalize globally. Additionally, access to AI technology may vary across healthcare settings, potentially exacerbating inequities in care delivery.

The future application of AI in thyroid cancer represents an exciting area of ongoing development. Emerging technologies include autonomous robotic ultrasound systems and AI-based malignancy risk stratification tools (9). AI is also being investigated intraoperatively to assist in identifying the recurrent laryngeal nerve and parathyroid glands, potentially reducing surgical morbidity. Furthermore, integrated computer models that synthesize imaging, pathology, molecular data, and clinical variables may surpass current nomograms by providing dynamic, individualized risk prediction. AI may also accelerate clinical trials in thyroid disease by efficiently analyzing comprehensive patient datasets and supporting shared decision-making.

In conclusion, although the use of AI in clinical practice remains in its infancy, it holds considerable promise. Current algorithms demonstrate variable effectiveness and are limited by concerns regarding generalizability, transparency, and accountability. Nevertheless, as with many other fields in medicine, AI is likely to play an increasingly important role in the diagnosis and management of thyroid pathology and may soon become an integral component of the clinician’s diagnostic and therapeutic armamentarium.

Acknowledgments

None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Annals of Thyroid. The article did not undergo external peer review.

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://aot.amegroups.com/article/view/10.21037/aot-2026-1-0013/coif). The authors have 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. Alhejaily AG. Artificial intelligence in healthcare Biomed Rep 2025;22:11. (Review). [Crossref] [PubMed]
2. Edström AB, Makouei F, Wennervaldt K, et al. Human-AI collaboration for ultrasound diagnosis of thyroid nodules: a clinical trial. Eur Arch Otorhinolaryngol 2025;282:3221-31. [Crossref] [PubMed]
3. Li Y, Liu Y, Xiao J, et al. Clinical value of artificial intelligence in thyroid ultrasound: a prospective study from the real world. Eur Radiol 2023;33:4513-23. [Crossref] [PubMed]
4. Chung SR, Baek JH, Lee MK, et al. Computer-Aided Diagnosis System for the Evaluation of Thyroid Nodules on Ultrasonography: Prospective Non-Inferiority Study according to the Experience Level of Radiologists. Korean J Radiol 2020;21:369-76. [Crossref] [PubMed]
5. Barczyński M, Stopa-Barczyńska M, Wojtczak B, et al. Clinical validation of S-Detect(TM) mode in semi-automated ultrasound classification of thyroid lesions in surgical office. Gland Surg 2020;9:S77-85. [Crossref] [PubMed]
6. Nair G, Vedula A, Johnson ET, et al. Combining Image Similarity and Predictive Artificial Intelligence Models to Decrease Subjectivity in Thyroid Nodule Diagnosis and Improve Malignancy Prediction. Endocr Pract 2024;30:1031-7. [Crossref] [PubMed]
7. Wu SH, Li MD, Tong WJ, et al. Adaptive Dual-Task Deep Learning for Automated Thyroid Cancer Triaging at Screening US. Radiol Artif Intell 2025;7:e240271. [Crossref] [PubMed]
8. Pozdeyev N, White SL, Bell CC, et al. Artificial Intelligence Applications in Thyroid Cancer Care. J Clin Endocrinol Metab 2026;111:316-24. [Crossref] [PubMed]
9. Su K, Liu J, Ren X, et al. A fully autonomous robotic ultrasound system for thyroid scanning. Nat Commun 2024;15:4004. [Crossref] [PubMed]