Artificial Intelligence in Thyroid Nodules and Cancer: Clinical Validation and Real-World Performance

Abstract

Thyroid nodules are detected in a large proportion of adults undergoing high-resolution ultrasonography, yet only a minority harbor clinically significant cancer. The clinical problem is therefore not only cancer detection but calibrated risk stratification: avoiding delayed diagnosis of aggressive disease while limiting unnecessary biopsies, molecular testing and diagnostic surgery. Artificial intelligence (AI) has moved rapidly from experimental image classification to clinically-deployed decision support. This invited review synthesizes current evidence for AI applications in the evaluation and management of thyroid nodules and differentiated thyroid cancer, emphasizing ultrasound-based computer-aided diagnosis, indeterminate cytology, molecular integration, cytopathology and histopathology, lymph-node assessment, report quality control, surveillance and emerging multimodal large language models. Commercial and near-commercial systems including S-Detect, AmCAD-UT, Koios DS Thyroid, AIBx and newer deep-learning systems show that AI can improve consistency, support less experienced readers and, in selected settings, reduce low-yield fine-needle aspiration without unacceptable loss of sensitivity. A particularly important future role may be AI-enabled de-escalation, in which image-derived estimates of benignity help support surveillance when clinical, sonographic, cytologic or molecular risk signals are concordantly low. However, performance varies by case mix, cancer prevalence, scanner platform, operator experience, geographic cohort, reference standard and whether the model is used as a stand-alone classifier or second reader. The strongest evidence supports AI as an adjunct to standardized ultrasound risk stratification and shared decision-making, not as a replacement for expert clinical judgment. Future progress will depend on prospective multicenter validation, transparent reporting, local calibration, workflow design, regulation, post-market surveillance and assessment of patient-centered outcomes.