ORF40-05 - Can Molecular Alterations in Thyroid Cancer Be Predicted from Routine Histology using Artificial Intelligence? A Systematic Review

Background
The rising incidence of low-risk thyroid cancer has led to increased detection without mortality reduction, raising concerns of overtreatment. Improved risk stratification is needed, with molecular status prediction playing a key role. However, molecular testing remains limited by cost, availability, and inconsistent use. Advances in artificial intelligence (AI) applied to digitized histopathology suggest that molecular alterations, including BRAF/RAS, may be inferred from routine hematoxylin–eosin (H&E) stained slides as digital biomarkers. Despite growing interest, clinical readiness of AI-based molecular prediction remains uncertain.

Methods
Following PRISMA 2020 guidelines and a registered PROSPERO protocol, we searched six databases from 2000 to September 2025. We included original studies evaluating AI models applied to digital whole-slide images of thyroid cancer. All task categories were reviewed, with a prioritized quantitative meta-analysis of molecular prediction studies. Risk of bias was assessed using adapted QUADAS-2 and PROBAST-AI tools.

Results
We screened 1,015 records and identified 11 studies evaluating AI-based molecular status prediction from thyroid histopathology (N = 13,174 patients), all using retrospective designs. Six thyroid-specific studies predicting BRAF V600E were sufficiently homogeneous for meta-analysis, yielding a pooled AUC of 0.91 (95% CI, 0.88–0.94). Across non-BRAF molecular targets, individual model performance was consistently high (reported AUCs 0.86–0.94), including models predicting RAS-related phenotypes, gene fusions, methylation patterns, and immune markers; however, heterogeneity in clinical targets and outcome definitions precluded pooling beyond BRAF. Only 3 of 11 studies incorporated independent external validation, including models predicting BRAF V600E, combined BRAF–RAS gene expression signatures, and panRAS mutations, using cohorts distinct from training datasets, with sustained discrimination (reported AUCs 0.88–0.94). The remaining studies relied on internal validation or public datasets, limiting assessment of real-world performance. Risk-of-bias assessment identified high risk in participant selection and analysis domains, mainly due to retrospective designs, non-representative cohorts, and limited external validation.

Conclusions
AI-based molecular prediction from routine thyroid histopathology shows strong performance, particularly for BRAF V600E, including externally validated models. However, evidence remains largely retrospective, and clinical readiness is limited by scarce independent external validation. Prospective, multi-institutional validation is needed before clinical implementation as an adjunct to molecular testing, while other AI-based clinical tasks identified here warrant future evaluation.