Accurate Detection Of Pulmonary Nodules Based On CT Imaging And Deep Learning

Focus detection is an important part of medical image analysis.However,traditional medical image analysis relies on doctors' understanding and recognition of images,which is a kind of image content interpretation based on prior knowledge.With the increasing of medical image data,data-driven new image analysis methods will gradually replace traditional methods.With the continuous development of machine learning and in-depth learning,lesion detection in medical imaging will usher in a new direction of development.Classification,detection and segmentation based on in-deph classification,detection and segmentation have formed a leading trend in image field,which proves its effectiveness and efficiency.As a kind of image,medical image has been highly evaluated in lesion detection.Therefore,lesion detection,based on in-deph learning,is an important research direction.In this paper,the detection of pulmonary nodules as the research object,aiming at the situation that pulmonary nodules have many low-level features and few semantic features,is a small target detection problem.At the same time,aiming at the problems of ambiguous boundary information,unclear classification boundaries and so on,this paper aims to develop a new algorithm for detecting and locating pulmonary nodules,so as to improve the accuracy of localization and reduce the false positive rate of detection.Therefore,this paper mainly makes the following improvements in the detection of pulmonary nodules:(1)Image preprocessing and region of interest processing are important links in medical image analysis.As a data-intensive algorithm,in-depth learning has a huge demand for data.Although there are some image data in public datasets,it is difficult to directly use as training data for in-depth learning because of the different ways of data formation.Therefore,this paper preprocesses the pulmonary nodule images,including pre-segmentation,cutting,labeling and data expansion.(2)Aiming at the low feature reuse rate of U-net convolution network,this paperintroduces extended convolution and dense connection network to improve the flow speed of feature information between input layer and output layer,To detect suspected nodules in lung images and remove false positive rate in suspected nodules.(3)Aiming at the lost characteristics of traditional convolution pooling operation,this paper introduces global average pooling to replace pooling operation when using deep learning model,and verifies its effectiveness in this problem.(4)Aiming at the problem of inaccurate location and high false positive rate of pulmonary nodules,this paper uses multi-model fusion to extract the boundary and regression frame of pulmonary nodules accurately,and designs a method of multi-layer image superimposed convolution to remove false positive pulmonary nodules by using the three-dimensional continuity between images.Through the above improvements,the detection accuracy can be improved by about 8%on average and the false positive rate can be reduced by about 1.7%.At the same time,the proposed method can provide clearer boundaries and more reliable candidate nodules in the localization of pulmonary nodules.The validity and scientificity of the proposed algorithm are verified by repeated and comparative experiments.