Gcan: Graph Convolutional Attention Network for Intraoperative Classification of Follicular Thyroid Neoplasms Using Whole Slide Image

Frozen section in intraoperative pathological diagnosis (IOPD) to diagnose a mass during surgery quickly and further determine the appropriate resection strategy during surgery. It is difficult to differentiate between follicular thyroid adenoma (FTA) and follicular thyroid carcinoma (FTC) because they share common clinical characteristics. Previous methods consider only patch information limiting their performance. We employed whole-slide images (WSIs) of thyroid frozen sections in a deep-learning system, introducing global information. The system consists of (i) pre-processing part, a joint of cancerous area segmentation, patch normalization, and channel separation that mitigates the effects of staining discrepancy on hematoxylin and eosin (H&E)-stained slides; and (ii) our proposed gatedattention network to aggregate patch-level information to slidelevel results. We managed to collect the largest data set of Follicular Thyroid Neoplasms WSI to the best of our knowledge. GCAN attains superior classification performance, reporting AUC scores of 0.972 and 0.912 on internal and external cohorts, respectively, while preserving strong generalization capabilities in the classification task of WSIs for FTA and FTC. Additionally, the model improved the accuracy of pathologists, potentially providing a significant auxiliary role for pathologists in diagnosing thyroid tumors.