Machine Learning for diagnosis of malignant thyroid nodules based on thyroid ultrasound: Systematic review and meta-analysis of studies with external datasets.

Introduction: Optimizing the diagnostic approach to thyroid nodules remains a crucial challenge. Ultrasound-based risk stratification systems such as EU-TIRADS have shown reasonable sensitivity and specificity. Therefore, we conducted a systematic review and meta-analysis to assess the diagnostic performance of Artificial Intelligence (AI) models in differentiating benign from malignant thyroid nodules on ultrasound data. Methods: A comprehensive search of PubMed/MEDLINE, Scopus, and Web of Science was performed up to January 1, 2025. Eligible studies included patients with thyroid nodules undergoing ultrasound, where AI-based models were validated against cytological or histological findings. The AI algorithms were developed using different types of ultrasound-derived data, including B-mode images, radiomics features. Pooled sensitivity, specificity, and area under the curve (AUC) were estimated using a hierarchical summary receiver operating characteristic (HSROC) model. Results: Twenty-seven studies comprising 146,332 patients and over 600,000 ultrasound images met inclusion criteria. Overall, pooled sensitivity was 87 % (95 % CI: 84-89 %) and specificity 83 % (95 % CI: 79-86 %). The summary operating point indicated a sensitivity of 88 % and specificity of 83 %, with an AUC of 91.9 % (95 % CI: 90.0-93.2 %). Although subgroup analysis suggested higher accuracy when cytology was used as the reference standard compared to histology, the mixed-effects meta-regression did not confirm a statistically significant association (p = 0.238 for sensitivity; p = 0.188 for specificity). Conclusion: AI-based algorithms show excellent diagnostic performance in distinguishing benign from malignant thyroid nodules, with robust validation across external datasets. These findings support the potential integration of AI into clinical thyroid nodule management, although further multicenter, non-Asian, and histology-based studies are warrantee. Systematic review registration: PROSPERO (CRD420251108149).