Study Of Thyroid Carcinoma Pathological Section Image Analysis Method Based On Deep Learning

Among the various types of cancer,thyroid carcinoma is the most common of endocrine malignancy and one of the cancers with the fastest increase in recent years.Although the cure rate of thyroid cancer is relatively high,its lymphatic metastasis rate is as high as 50%,and papillary thyroid cancer accounts for more than 80%of all thyroid cancers.Therefore,the prevention and early diagnosis of papillary thyroid cancer are of enormous significance for preventing and treating the thyroid cancer.Pathological diagnosis is the "gold standard"for accurate diagnosis,treatment and prevention of thyroid cancer.However,due to medical resource problems,how to use the increasingly developed computer technology to assist pathologists in pathological diagnosis is extremely urgent.This paper focuses on the most basic and important tasks of thyroid pathological image classification and cell nucleus segmentation and classification,and carries out research on the image analysis method of thyroid cancer pathological sections based on deep learning.Aiming at the inaccurate feature extraction of the existing pathological image classification methods for papillary thyroid carcinoma pathological images and the low classification performance,we propose a pathological image classification method for papillary thyroid carcinoma based on capsule network.This method uses an improved capsule network and designs the pathological semantic feature extraction module,which can extract features that are more conducive to classification from the pathological slice images of thyroid cancer.The experimental results show that the presented method can extract semantic features of pathological images more accurately,and achieves higher classification accuracy on the PTC pathological image database.In order to solve the problem that the existing methods only use a single magnification for pathological image classification and the low classification accuracy of complex pathological images,we propose a method for PTC pathological image classification based on multi-magnification feature fusion.This method simulates the pathological analysis process of pathologists,and classifies the PTC pathological images by simultaneously using pathological images with multi-magnifications for feature fusion.At the same time,a channel attention ASPP(CA-ASPP)module is designed to extract semantic features of pathological images with different magnifications.And we design a metric named FPN-score,and measure the classification accuracy,FNR and FPR at same time.Experiments show that our method can effectively improve accuracy for PTC pathological images classification.Since pathological image classification can only determine whether there is cancer area,further pathological diagnosis requires nucleus segmentation and classification.In view of the fact that the existing nucleus segmentation and classification methods based on deep learning are poor in the classification and segmentation of adjacent nuclei and the nucleus boundary segmentation is inaccurate,a method for PTC pathological images nucleus segmentation and classification based on adjacent attention is proposed.This method uses the pyramid pooling module to extract the semantic information of different regions in multi-scale thyroid cancer pathological images,and proposes the block pooling module to learn the semantic relationship between adjacent nuclei which improve the classification and segmentation accuracy of adjacent nuclei.Experiments show that this method can effectively improve the accuracy of nucleus segmentation and classification.