Pathological Imaging Study Based On Deep Convolutional Neural Network

Pathological diagnosis,as the "gold standard"for medical diagnosis,is one of the important means for the diagnosis of many diseases.However,pathological diagnosis has problems such as strong subjectivity,high rate of clinical misdiagnosis and large gaps in pathologists.In recent years,with the development of deep learning and Whole Silde Images(WSIs)acquisition technology,artificial intelligence technology has gradually entered the field of pathological imaging.This not only makes the pathological diagnosis more accurate and obje.ctive,but also greatly reduces the burden on the pathologist.The research goal of the thesis is to explore the recent progress of deep learning algorithms in the field of pathological imaging in recent years,and to conduct detailed research,analysis and comparison on the three tasks of gastric cancer lesion segmentation,breast tumor cell score and multi-organ organ cell nuclear segmentation.The main work of the research on the advantages and disadvantages of various methods includes the following aspects:1.Aiming at the pathological data of crude gastric cancer,a double-input convolutional neural network was proposed.The classification result of the patch was converted into the segmentation result of the image by a voting splicing algorithm,so as to achieve the purpose of tumor segmentation of gastric cancer pathology.2.Aiming at the task of breast tumor cell scoring,a collaborative deep convolutional network was proposed to evaluate the cell characteristics of breast pathological sections,and the regression performance of the network model was improved by paired input data and synergistic errors.3.Aiming at the segmentation task of multinuclear organ nucleus instance,a double output convolutional neural network model based on cavity convolution is proposed.The cell nucleus and its edges are segmented,and then the marker-based watershed algorithm is used for post-processing.Example segmentation of multiple tissue organ nuclei.Finally,the paper validates the experimental model of the proposed algorithm in the test set of the corresponding task.The experimental results show the effectiveness and robustness of the algorithm.The series of fully automatic analysis algorithms proposed in this paper provide reference for the application of subsequent deep learning in the field of digital pathology,and advance the process of computer technology to assist doctors in diagnosis,which has certain clinical value.