Abstract / Summary
Artificial intelligence (AI) for thyroid nodules should be evaluated against the decisions it changes, not diagnostic accuracy alone. This critical narrative review examines ultrasound, cytology, histology, and molecular models and distinguishes modality-specific prediction from implemented multimodal learning. The original PubMed/MEDLINE search through 10 August 2026 was supplemented on 25 September 2026 by targeted OpenAlex searches and checks of primary publications, reporting standards, and regulatory sources. Selected studies are compared by sample size, reference standard, acquisition setting, testing design, and performance with confidence intervals where available. We explain feature concatenation, intermediate representation learning, cross-attention, and decision-level fusion, and examine whether integration improves on the strongest component. Reported gains are inconsistent, and complete-case multimodal datasets may exclude the low-risk nodules most relevant to avoiding biopsy. High discrimination for histological malignancy or occult nodal disease does not demonstrate reduced overdiagnosis, fewer operations, or improved survival. A thyroid-specific translation roadmap therefore addresses active surveillance, partial verification and spectrum bias, Hashimoto thyroiditis and multinodular disease, cytology preparation shifts, category-specific thresholds, operator dependence, reporting frameworks, and regulatory change control. The priority is a calibrated, externally tested tool using information available at a defined clinical decision, followed by prospective evaluation of patient and workflow outcomes. A unified imaging–pathology–omics system and reliable AI prediction of surveillance progression remain research objectives rather than established clinical capabilities.