Research On Key Techniques Of Computer Aided Segmentation And Recognition Of Pathological Images Of Lung Adenocarcinoma

Lung cancer is a common malignant tumor.It is also the cancer with the highest morbidity and mortality at home and abroad.Lung adenocarcinoma is one of the common lung tissue lesions in lung cancer.However,due to the high complexity of pathological images,doctors need to repeatedly read a large number of pictures to finally give a medical diagnosis,which makes the burden of reading a doctor significantly heavier.Therefore,the automatic recognition and precise positioning of lung adenocarcinoma pathological images with the help of computers is of great significance for improving the diagnosis rate.This thesis focuses on the research of image registration technology for sequence pathological slices and image segmentation technology based on deep learning during the pathological examination of lung adenocarcinoma,in order to achieve the purpose of computer-aided identification and detection.The main research contents of the thesis are as follows:1.An improved image registration algorithm R-SIFT is proposed.The algorithm introduces an enhanced gradient to extract image feature points.The dual matching process of initial matching and fine matching is used to effectively increase the number of feature matching point pairs while reducing false matching point pairs.The pathological image registration test of lung adenocarcinoma was analyzed and compared with SIFT + RANSAC and SIFT + FSC algorithms.The number of feature matching points and the root mean square error were significantly improved,which verified that the proposed registration algorithm had better accuracy and robustness.2.An image semantic segmentation model based on deep learning is used to detect lung adenocarcinoma pathological images.Based on the existing Deeplab V3 + network model,an improved L-Deeplab network is proposed for better pathological image segmentation.The network mainly adopts an encoding-decoding structure,in which the encoding structure optimizes the backbone network Xception and the hollow space pyramid pooling module to extract features of different scales,and then obtains depth features.The decoding structure decodes the extracted features,restores the image feature information by fusing low-level semantic information and high-level semantic information,and finally achieves the purpose of image segmentation.3.The improved L-Deep Lab network was applied to the image processing of lung adenocarcinoma pathological slices,and the image segmentation of lung adenocarcinoma pathological slices was realized,which verified the feasibility and robustness of the network.Compared with the fully convolutional neural network,U-Net network,Deep Labv3 + network model,the experimental results show that: in the segmentation task of lung adenocarcinoma pathological images,the global pixel accuracy of the L-Deeplab network reaches 91.48% and the Mean Intersection over Union(MIo U)reaches 84%.The performance of the model segmentation performance and detection speed are good,and it can realize the segmentation recognition of lung adenocarcinoma pathological images,provide reliable pathological information for physicians,and assist physicians in achieving efficient diagnosis.