Research On Detection And Classification Of Small Pulmonary Nodules Based On Convolution Neural Network

Lung nodule detection is the first step in lung cancer screening,and lung nodule detection in the early stage of lung cancer is very important for early treatment of lung cancer and improving survival rate of lung cancer patients.CT is the main means of early lung cancer screening,in patients with lung cancer early CT images,pulmonary nodules tiny volume,blood vessels,windpipe,bubble and so on organization,makes the lung small nodules detected faces huge challenges.With the explosive growth of CT image data and the development of computer image processing technology,computer-aided detection system has become a powerful tool in the work of the radiologist.The rapid growth of CT image data in lung provides an opportunity for the application of deep learning techniques in medical imaging.For small size of lung tiny nodule images and difficult to accurately detect the problem,based on public data sets LIDC constitute a huge amounts of training set,this paper puts forward multiple-input convolution neural network and convolution deconvolution neural network(CDNN)for detecting lung small nodule in lung CT image,fast and efficient detection of lung nodules,provide auxiliary reference basis for physicians in the early screening.The specific research content of this paper is as follows:(1)Multi-input convolution neural network detection method research.As large range of the scale in the CT images of pulmonary nodules,using convolution neural network in lung nodules in the process of image detection,image preprocessing will reduce the data information in input stage.For improve the convolutional neural network feature extraction ability,we design a multiple input convolution neural network method.This method is mainly divided into offline learning and online detection.For offline learning phase,use the build data set a variety of scale of pulmonary nodule image as the training images,through simple data transform and scale transform as a multi-input convolution of the neural network input,and then through joint training learning mechanism of the network.For online detection phase,combined with prior knowledge,in the original CT images using two-dimensional Gaussian function and Canny edge detection algorithm is combined to generate the sampling points,extract a variety of scale detection area,detecting and CT images.In order to evaluate multiple input convolution of the neural network classification performance,in a public data set of reserved contrast experiment on the test set for multiple sets of parameters,the optimal results in the coverage rate was 85.51%,and the detection accuracy up to 78.4%.(2)Study on Detection Method of Tiny Pulmonary Nodule Based on Convolution Deconvolution Neural Network.The tiny lung nodule images difficulty design features filter in CT.As deep learning approach is a kind of based on data driven,end-to-end mapping method of study,thus put forward a kind of convolution deconvolution neural network tiny lung nodules detection method.Our method,convolution deconvolution neural network,is the first to use the contrast of unsupervised divergence algorithm to learn feature reverse remodeling,and then use stochastic gradient descent algorithm of supervised classification learning of nodules-non nodules in offline learning part using lung small nodule images,online testing phase with multi-input convolutional neural network is adopted at the same area suggest structure,by rotating and scale transformation to achieve a variety of original data input,the joint detection by using parallel network method lung small nodule area.Group through many comparative experiments,and a variety of machine learning methods validation deconvolution neural network in the lungs of tiny nodules convolution feature extraction is effective,through a public data set aside on the test set validation testing accuracy.(3)LIDC data set visualization tool.The deep learning method has a large demand for data volume.In the process of research,the statistics and use of public data sets become part of the work.Are introduced in detail in this paper the LIDC data visualization tools,such image merging algorithm,clear in the annotations in the process by the merger of the rules,the data centralized data distribution statistics,briefly analysis on the public data sets can continue to carry out research work.