Research On Tissue Segmentation Of Lung Cancer Pathological Images Based On Deep Convolutional Neural Network

Among malignant tumors,lung cancer has the highest incidence and mortality rate in our country.The life and health of our people is severely affected by this cancer.It is important for doctors to choose appropriate treatment plan through accurate diagnosis to control the development of lung cancer and cure patients.Pathological diagnosis is the "Gold Standard" of lung cancer diagnosis.Pathologists can obtain a wealth of tumor microenvironment information,such as tumor tissue,necrotic area and other important information,by observing pathological sections under a microscope,so as to make diagnosis of patients’ conditions.With the development of digital pathological sections,automatic segmentation and quantitative analysis of lung cancer pathological tissue data has become a key step in accurate treatment.Recently,various studies using deep convolutional networks have been proposed on pathological tissue data sets.However,due to the high heterogeneity of lung cancer pathological tissues,deep convolutional networks are difficult to identify lung cancer tissue types.In addition,a lots of annotated data are required to train the deep convolutional networks,but it is difficult to annotate lung cancer pathological tissue data sets,which seriously affects the experimental progress of researchers.In order to realize the automation of tissue segmentation and quantitative analysis,as well as ease the pressure of data annotation by doctors,this paper explores and researches from the following two directions.(1)Bilinear pathological image segmentation of lung cancer based on attention mechanism: A simple and accurate tissue segmentation algorithm is designed in this paper for the problems of high heterogeneity of lung cancer pathological tissues,large intra-class variability and small inter-class variability.In order to deeply capture the subtle features of pathological tissue types.Bilinear module is introduced in this paper to fully explore the semantic relations among pathological tissues,and combined with the attention mechanism.The deep convolutional network focuses its learning on the local invariant features of tissue types.And the spatial mapping of data segmentation results is completed through the tissue segmentation mapping algorithm,which can complete the automation of tissue segmentation and reduce the redundant calculation.The proposed method is verified on lung cancer data set,and the accuracy rate of tissue type recognition in lung cancer pathological tissue data set is 93.94%.Experiments are also conducted on a publicly available data set of pathological tissues of colorectal cancer,and the accuracy rate of colorectal cancer tissue classification is 98.23%.The experimental results prove the robustness of the proposed method.In addition,the interpretability of our proposed method is demonstrated by Grad-CAM.(2)Tissue recognition in semi-supervised pathological images based on non-local mechanism: A fully supervised deep convolutional network relies on large amounts of annotated data.However,pathological tissue images are difficult to annotate due to the complex and subtle characteristics of data.To solve this problem,this paper proposes a semi-supervised pathological tissue recognition algorithm to alleviate the dependence of the algorithm on labeled data.According to the learning state of each tissue type in the training process of the deep convolutional network,the threshold of pseudo-label is dynamically adjusted to improve the use of unmarked data by the network.In addition,in view of the high heterogeneity of lung cancer pathological tissue data,this paper introduces a non-local module to enhance the ability of deep convolutional network to capture pathological tissue features,focusing on the internal correlation of tissue types.The consistency regularization mechanism is used to guide the training of deep convolutional networks,and a handful of labeled data is used to complete the task of tissue type recognition.At the same time,experiments are carried out on lung cancer pathological tissue data sets and colorectal cancer data sets.In the comparison experiments with different labeled data,the tissue type identification accuracy of the semi-supervised method proposed in this paper is the first.