Research On Classification Algorithm For Thyroid Frozen Digital Pathological Images

In recent years,malignant thyroid nodules,also called thyroid cancer,have become one of the fastest-growing malignant tumors.When the malignant degree of the thyroid nodule tends to be high,pathologists need to conduct an intraoperative freezing pathological diagnosis to help the clinical doctor determine the next surgical plan during thyroid surgery.However,the low quality of thyroid frozen slides and the need for rapid diagnosis make frozen pathological diagnosis difficult.At the same time,the shortage of pathologists and the increasing demand for intraoperative frozen pathological diagnoses lead to heavy work for pathologists.Facing such difficult and heavy work,pathologists will inevitably miss and misdiagnose thyroid cancer,which makes the use of computer technology to assist pathologists in diagnosing frozen thyroid slides necessary.Thus,this paper focused on the challenging problem called thyroid frozen digital pathological image classification and carried out the following studies:(1)Considering that the methods supervised by patch-level labels have the design flaw of relying on hand-crafted rules or features,this paper proposes a dual encoder network.The dual encoder network uses a convolutional neural network as the slide-level classifier to automatically extract features from the heatmap generated by the patch-level classifier.Experiments show that the network can extract better classification descriptors and achieve96.03%,100%,94.01%,and 96.91% in Accuracy,Precision,Recall,and F1-Score,respectively.(2)Considering that the methods supervised by slide-level labels have the design flaw of ignoring contextual relations,this paper proposes a context-aware one-stage network.It uses a recurrent neural network to model the contextual relations among all image patches.It also uses a feature attention module to weigh each image patch’s importance.Experiments show that this method outperforms other methods supervised by slide-level labels and can achieve 92.06%,97.42%,90.42%,and 93.79% in Accuracy,Precision,Recall,and F1-Score,respectively.(3)Aiming at the misdiagnosis problem,this paper proposes a false negative penalty loss function to reduce the misdiagnosis rate.The core idea is punishing false negative samples,that is,punishing positive samples that are misclassified as negative ones.Experiments show that this loss function can improve the recall value of the dual encoder network by4.19% and the recall value of the context-aware one-stage network by 5.99%.