Thyroid nodules are common and sonographic interpretation is subject to interobserver variability. This systematic review aimed to evaluate the diagnostic accuracy, clinical utility, and workflow integration of AI-assisted sonography for thyroid nodule assessment. The review also examined whether AI tools can improve diagnostic consistency, reduce unnecessary fine-needle aspiration (FNA) procedures, and support clinical decision-making.
Database searches identified studies published between January 2018 and June 2025. Eligible studies included those that compared AI-based models for thyroid sonography against histopathology or FNA as reference standards. The review followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 criteria for data selection and extraction.
Extracted variables comprised AI architecture, sonography mode or technique, study design, clinical impact measures (for example, FNA reduction and time savings), and diagnostic metrics such as sensitivity, specificity, and area under the curve (AUC). Risk-of-bias assessments were performed for included studies.
A total of 30 studies were included in the review. The most commonly reported AI approach across those studies was convolutional neural networks. The included reports covered experimental models through to systems approaching clinical implementability. The review summarized how AI models were developed and externally validated in some instances, and how they were paired with established clinical frameworks such as the American College of Radiology Thyroid Imaging and Reporting Data System (TI-RADS).
Across the included studies, reported diagnostic performance varied but demonstrated generally high efficacy. Area under the curve (AUC) values ranged from 0.78 to 0.97. Reported sensitivity values ranged from 75.6% to 94.0%, and specificity values ranged from 70.4% to 90.7%. These metrics indicate that many AI models achieved substantial discriminatory ability for differentiating benign from malignant thyroid nodules when compared with histopathology or FNA reference standards.
Several studies quantified clinical advantages beyond diagnostic metrics. The review reported reductions in the number of FNAs performed, with study-level estimates ranging from a 10% to 45% decrease in FNAs. Some studies also reported time savings in interpretation, with values between 2.5 and 3.1 minutes saved per case. These findings suggest potential workflow benefits and reduced procedural burden when AI tools are deployed as adjuncts to sonographic assessment.
Risk-of-bias assessment categorized 53.3% of the included studies as low risk, 33.3% as moderate risk, and 13.3% as high risk. The principal sources of higher risk were retrospective study designs and small sample sizes in some reports. The distribution of study quality highlights the need for more prospective and adequately powered evaluations.
The review emphasizes that AI-assisted sonography can improve diagnostic consistency and that performance is particularly compelling when AI systems are combined with TI-RADS and subjected to external validation. The authors map the progression from prototype experimental models to systems that could be implemented clinically. They argue that standardization, prospective multicenter evaluation, and efforts to make AI outputs explainable are necessary steps to ensure safe and reproducible use in routine thyroid imaging.
This systematic review concludes that AI-aided sonography demonstrates high diagnostic efficacy and measurable clinical benefits, including reductions in FNAs and modest per-case time savings in some studies. The authors call for:
The review also links these improvements to broader goals: supporting Sustainable Development Goal 3 (good health and well-being) by promoting higher-quality diagnostic care and SDG 9 (industry, innovation, and infrastructure) by advancing AI in medical imaging.
The source identified limitations among included studies, notably retrospective designs and small sample sizes that contributed to moderate or high risk-of-bias ratings in a subset of reports. Specifics beyond these general concerns (such as exact study-level limitations or individual study protocols) were not detailed in the abstract.
Overall, the body of evidence from 2018–2025 included in this review supports the diagnostic promise of AI-assisted sonography for thyroid nodules while underscoring the need for prospective, standardized, and explainable approaches before widespread clinical implementation.