Purpose To explore the artificial intelligence(AI)-assisted diagnosis system of thyroid cancer based on deep transfer learning and evaluate its clinical application value.Methods The HE sections of 682 cases thyroid disease patients(including benign lesions,papillary carcinoma,follicular carci-noma,medullary carcinoma and undifferentiated carcinoma)in the Pathology Department of the Renmin Hospital of Wuhan Uni-versity were collected,scanned into digital sections,divided into training sets and internal test sets according to the ratio of 8 ∶ 2,and the training sets were labeled at the pixel level by patholo-gists.The thyroid cancer classification model was established u-sing VGG image classification algorithm model.In the process of model training,the parameters of the breast cancer region recog-nition model were taken as the initial values,and the parameters of the thyroid cancer region recognition model were optimized through the transfer learning strategy.Then the test set and 633 intraoperative frozen HE section images of thyroid disease in Jianli County People's Hospital,Jingzhou City,Hubei Province wereused as the external test set to evaluate the performance of the established AI-assisted diagnostic model.Results In the internal test set,without the use of the breast cancer region rec-ognition model transfer learning,the accuracy of the AI-assisted diagnosis model was 0.882,and the area under the Receiver op-erating characteristic(AUC)valuewas0.938;However,inthe use of the Transfer learning model,the accuracy of the AI-assis-ted diagnosis model for was 0.926,and the AUC value was 0.956.In the external test set,the accuracy of the zero learning model was 0.872,the AUC value was 0.915,and the accuracy of the Transfer learning model was 0.905,the AUC value was 0.930.Conclusion The AI-assisted diagnosis method for thy-roid cancer established in this study has good accuracy and gen-eralization.With the continuous development of pathological AI research,transfer learning can help improve the performance and generalization ability of the model,and improve the accura-cy of the diagnostic model.