Artificial intelligence may help improve management of indeterminate thyroid nodules by reducing variability in cytologic interpretation and integrating ultrasound, molecular, and clinical data into more consistent risk assessments, according to a narrative review published in Frontiers in Endocrinology. 

Researchers from the University of Messina and Papardo Hospital in Messina, Italy found that AI shows promise to support decision-making throughout the thyroid nodule diagnostic pathway, particularly by improving standardization, reproducibility, and multimodal risk assessment. However, they noted that prospective studies, standardized digital cytology workflows, and external validation are needed before AI can be adopted routinely in clinical practice, and that it should complement—not replace—cytopathologists or molecular diagnostics.  

Indeterminate thyroid cytology (the diagnostic "gray zone") remains one of the greatest challenges in thyroid nodule management because cytology cannot reliably distinguish many benign from malignant follicular-patterned lesions.  Although molecular testing has improved preoperative risk stratification, uncertainty surrounding Bethesda III and Bethesda IV nodules continues to result in repeat biopsies, molecular testing, specialist consultations, and diagnostic surgery for many patients with ultimately benign disease. 

To address these challenges, the authors suggested that AI may be most valuable for improving reproducibility, reducing subjectivity, and standardizing risk stratification across institutions. 

They described AI as a decision-support tool across the thyroid nodule diagnostic pathway, with potential applications in ultrasound-guided biopsy selection, AI-assisted interpretation of digitized cytology, and multimodal integration of cytologic, molecular, imaging, and clinical data to support decisions between surveillance and surgery.  

Among the most promising applications is AI-assisted whole-slide imaging and computational pathology, in which deep learning models analyze digitized pathology slides to identify cytomorphologic features associated with indeterminate thyroid cytology—including microfollicular architecture, oncocytic change, nuclear atypia, colloid-rich backgrounds, and cellularity patterns. The authors suggested these systems could improve quantitative morphologic assessment and reduce interobserver variability within Bethesda III and IV categories. 

The authors also highlighted AI-assisted identification of lesions within the RAS-like spectrum, including noninvasive follicular thyroid neoplasm with papillary-like nuclear features, where subtle morphologic differences frequently contribute to diagnostic disagreement.  

Additional proposed workflow applications include automated adequacy assessment, quality-control monitoring, slide screening, region-of-interest detection, and second-reader assistance for diagnostically challenging cases.  

No single diagnostic modality captures the complete biologic signal of an indeterminate thyroid nodule. Instead, the greatest opportunity may lie in multimodal AI systems that integrate ultrasound, cytology, molecular testing, and clinical variables into unified predictive models. The authors emphasized that the ultimate goal is not simply to improve diagnostic performance but to reduce unnecessary biopsies and surgery while preserving safe surveillance for patients with low-risk disease. However, they noted that most supporting evidence remains retrospective or early stage and requires validation before routine clinical use. 

The authors summarized several previously published AI models for indeterminate thyroid nodules. Among multimodal approaches, one deep learning model combining ultrasound and molecular testing achieved 95% sensitivity and 66% specificity, while another integrating AI-assisted ultrasound, cytology, and clinical data achieved 90% overall accuracy, 90% sensitivity, and 91% specificity for Bethesda III-IV nodules. An AI-based decision-support system for ultrasound improved positive predictive value from 33% to 55% while maintaining a negative predictive value of 94%, according to the cited studies. 

The review authors also discussed expanding AI applications in proteomics and spatial profiling. AI-based proteomic classifiers and protein-expression signatures have shown potential to improve classification of benign vs malignant thyroid nodules, while digital analysis of immunocytochemical biomarkers could supplement diagnosis in settings where molecular testing is unavailable.

The authors noted that AI systems intended for clinical implementation should undergo multicenter external validation, provide calibrated risk estimates aligned with local malignancy risk, demonstrate prospective clinical benefit, and include continuous post-deployment monitoring before widespread implementation. 

They concluded that future AI studies should evaluate clinically meaningful outcomes rather than focusing solely on discrimination metrics. Potential endpoints include reductions in unnecessary biopsies and diagnostic surgery, optimization of molecular testing, workflow efficiency, cost-effectiveness, and preservation of safe surveillance strategies for patients with low-risk disease. AI may also serve as an educational second reader for trainees if safeguards against automation bias are incorporated. 

However, the authors identified several barriers to clinical implementation, including variability in slide preparation, staining protocols, scanner platforms, digitization methods, and molecular testing availability. Others include reduced model performance across institutions, selection bias, limited external validation, workflow integration challenges, and quality-control requirements.  

"Ultimately, successful clinical implementation will depend not only on algorithmic performance but also on transparency, regulatory oversight, interoperability with existing pathology workflows, and continuous quality-control monitoring," the authors wrote. 

No commercial or financial conflicts of interest were reported. Valeria Zuccalà, MD, disclosed serving as an editorial board member for Frontiers in Endocrinology at the time of submission, although the journal stated this did not influence peer review or editorial decisions.