Interpretable AI-assisted diagnosis of papillary thyroid cancer cytopathology using graph neural networks and knowledge graphs

Abstract

Abstract This study presents an interpretable AI-assisted diagnostic approach for papillary thyroid carcinoma (PTC) cytopathology by combining graph neural networks (GNNs) with knowledge graphs (KGs). Routine cytology smears from 281 PTC cases were scanned, labeled, and processed using the Cascade RCNN model to detect pathological cell features, including 45,680 ground-glass nuclei, 712 nuclear grooves, and 116 intranuclear inclusions. By integrating GNNs, the model achieved a mean intersection over union (mIoU) of 56.14% and a mean average precision (mAP) of 0.87. The GINet model further improved classification accuracy to 88.84%. Our approach also incorporates a clinical decision support system (CDSS) for querying KGs, providing explainable diagnostic outputs. This method offers an interpretable and reliable AI tool for PTC diagnosis, enhancing the transparency of AI-assisted pathology systems.