Artificial intelligence-based systems as auxiliary tools for thyroid nodule malignancy detection on ultrasound: a systematic review and meta-analysis of diagnostic accuracy.

BACKGROUND: Artificial intelligence (AI)-based systems are valuable tools in thyroid nodule diagnosis, but clinical controversy remains regarding diagnostic inconsistency and specificity. This meta-analysis evaluated the diagnostic accuracy of AI-assisted ultrasound systems for detecting malignant thyroid nodules compared to radiologists and risk stratification systems. METHODS: Guided by the PICOS (Population, Intervention, Comparison, Outcome, Study Design) framework, a systematic search across PubMed, Semantic Scholar, and OpenAlex identified 69 studies using histopathological or cytological reference standards. Quality and Risk of Bias were assessed via an AI-tailored QUADAS-2 tool. Pooled sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were calculated using bivariate random-effects meta-analysis. RESULTS: The 69 studies evaluated >350,000 nodules. AI-based systems demonstrated high diagnostic accuracy: pooled sensitivity 89% [95% confidence interval (CI): 87-91%], specificity 84% (95% CI: 80-88%), and AUC 0.93 (95% CI: 0.91-0.95). In direct comparisons, AI systems achieved comparable or superior performance to experienced radiologists (AUC 0.90-0.95 vs. 0.76-0.86). AI significantly improved specificity over individual radiologists and systems like ACR TI-RADS. AI-optimized TI-RADS achieved higher specificity (70.2% vs. 49.2%, P<0.001) than standard ACR TI-RADS with similar sensitivity. Consequently, AI-assisted strategies reduced unnecessary fine-needle aspirations (FNAs) by up to 26.7 percentage points. Performance peaked for nodules <20 mm and in externally validated, multicenter settings. AI provided the greatest incremental benefit for less experienced practitioners, markedly improving interobserver agreement. CONCLUSIONS: AI systems achieve high diagnostic accuracy for thyroid nodule malignancy, demonstrating performance comparable or superior to radiologists and risk stratification frameworks. As adjunctive tools, they safely reduce unnecessary biopsies and improve diagnostic consistency. Future clinical integration must prioritize multimodal data and radiomics to maximize utility.