Deep learning‐based prediction of cervical lymph node metastasis and genetic alterations from whole‐slide images of thyroid cancer frozen sections

Abstract Precise evaluation of cervical lymph node metastasis (CLNM) and genetic mutations (BRAF V600E /TERT promoter, TERTp) is pivotal for tailoring surgical and prognostic evaluation and adjuvant strategies in thyroid cancer (TC). Although current methods have limitations, we aim to develop deep learning (DL) models to predict CLNM and genetic mutations from TC frozen sections. We developed a DL framework using 2499 frozen‐section whole‐slide images from 2176 TC patients across five centers. The model was trained with a transfer learning‐based feature extractor and an attention‐based multiple instance learning (MIL) classifier, and validated on both internal and external cohorts. StyleGAN3‐based data augmentation was employed to tackle class imbalance for TERTp prediction, while interpretability was assessed via attention heatmaps and Leiden clustering. The CLNM prediction model achieved a patient‐level AUROC of 0.918 internally and 0.803–0.885 across three external validation datasets. For BRAF V600E prediction, AUROCs attained 0.814 internally and spanned 0.750––0.811 in external validation. In TERTp mutation prediction, GAN‐based augmentation increased the AUROC to 0.804 (internal) and 0.732 (external), up from 0.782 and 0.724, respectively. Attention maps visualized CLNM correlations with invasive tumor margins, while mutations localized to specific cellular morphology features. Our DL models accurately predict CLNM and genetic mutations from TC frozen sections, potentially reducing unnecessary procedures and providing a rapid alternative to traditional molecular testing.