Beyond genomics: artificial intelligence-powered diagnostics for indeterminate thyroid nodules-a systematic review...

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Thyroidology

Background:

• Indeterminate thyroid nodules (ITNs) present a diagnostic challenge in clinical thyroidology, impacting surgical decision-making.
• Genomic sequencing classifiers (GSCs) aid in ITN diagnosis but face barriers of cost and accessibility, creating an equity gap.
• Artificial intelligence (AI) tools are increasingly studied for thyroid ultrasonography (USG) classification, aiming to improve pre-operative diagnosis.

Purpose of the Study:

• To systematically review and analyze the current evidence on AI tools for diagnosing ITNs without relying on GSCs.
• To assess the diagnostic accuracy and clinical applicability of AI models in the pre-operative assessment of ITNs.
• To identify limitations and areas for future development in AI for ITN diagnosis.

Main Methods:

• A systematic literature search was conducted on PubMed, Google Scholar, and Scopus up to February 18, 2025.
• Included studies evaluated the diagnostic accuracy of AI for ITNs, with a focus on models not using GSC.
• A meta-analysis was performed on area under the curve (AUC) results from included AI models, adhering to Cochrane Collaboration guidelines.

Main Results:

• Seven studies presented 20 AI models, including radiological, natural language processing, and cytology-focused approaches.
• The pooled meta-analysis of 15 models yielded a combined AUC of 0.82, indicating moderate-to-good classification performance for machine learning (ML) and deep learning (DL) architectures.
• Substantial heterogeneity was noted, particularly in DL models (AUC=0.85), while ML models showed minimal heterogeneity (AUC=0.75). Meta-regression suggested potential publication bias or systematic differences in model design and validation.

Conclusions:

• AI demonstrates significant potential to aid clinical decision-making for ITNs, offering a promising alternative to GSCs.
• Current AI models, while performing adequately, are not yet suitable for widespread clinical implementation due to limitations in performance and lack of robust external validation.
• Further research and development are essential to enhance AI model robustness, generalizability, and clinical utility in the pre-operative diagnosis of thyroid nodules.