Study Of Thyroid Carcinoma Pathological Image Classification Based On Deep Learning

As the most common tumor of the endocrine system,thyroid cancer is also the most rapidly increasing malignant tumor all over the world.Papillary thyroid cancer is the most common type of thyroid cancer,and early diagnosis can make a high survival rate as high as 90%.However,the rate of lymphatic metastasis is high.Therefore,early diagnosis of papillary thyroid cancer(PTC)plays an important role in formulating appropriate treatment plans and preventing the deterioration of the disease.Among the multiple diagnostic methods of thyroid cancer,the pathological examination is the most sensitive and reliable diagnostic method.Facing the increasing number of pathological samples,how to use computer technology to assist the pathologist in diagnosis and reduce the burden on the pathologist is the key problem to be solved.This paper focuses on the pathological image classification,which is an important and extremely challenging subject,and carries out research on automatic classification of thyroid cancer pathological images based on deep learning.To address the problem that the existing pathological image classification methods are low in accuracy and only use a single magnification factor,we propose a multiple magnification factors synergetic method for PTC pathological image classification which imitates the diagnosis process of the pathologist and employs attention mechanism to obtain the discriminative regions.Then those discriminative regions are magnified and patched for further analysis.Finally,the classification result is obtained by fusing the classes of those image patches.At the same time,transfer learning is used to narrow the difference caused by H&E staining and improve the classification accuracy of pathological images.The experimental results demonstrate that the proposed method can make full use of the information for pathological images with different magnifying factors and obtain good performance on the established database.In order to avoid the problems that current deep-learning based classification methods require a large amount of labeled data while the annotation of pathological images is difficult to acquire,this paper constructs a classification method based on deep active learning for PTC pathological images classification.This method integrates deep learning and active learning into a framework,and uncertainty estimation is realized through the similarity of hash codes and the entropy of samples.The representativeness of samples is evaluated by convolution features extracted from CNN.Finally,the model is further fine-tuned by adding annotated samples in each iteration to improve the classification accuracy of the model.The experimental results show that the proposed method can effectively reduce the labeling cost and use fewer labeled samples to achieve good classification accuracy.In view of the complexity of deep classification networks,a method for PTC pathological image classification with integrated feature distillation was presented in this paper to assist pathologists to analyze and judge pathological images quickly.Based on the teacher-student network,this method guides the student network to learn deep features similar to the teacher network and promotes the lightweight student network to achieve the performance of the teacher network so that the model can be more easily deployed for practical application.Experiment results show that this method can maintain the classification effect of deep teacher networks while reducing memory usage and computation.