Research Of Deep Detection And Prediction For Pulmonary Nodule Based On CT Image Data

Among the causes of death from malignant tumors,lung cancer is the leading cause of death.About 75% of patients have been found to have been in the middle or late stage of lung cancer.Therefore,early specific screening for lung cancer is necessary.So it is urgently to realize early specific screening of lung cancer.The cure rate of early lung cancer patients can reach more than80%.On the basis of overcoming the problems of traditional x-ray imaging duplication and bone overlap,CT technique provides a high-definition CT image sequence with high resolution and contrast for doctors in the diagnosis of various organs and tissues of the human body.At the same time,the computer-aided diagnosis system makes use of the ability of high speed calculation and automatic processing of the computer to make images more accurate to reflect the actual state of illness and help doctors to diagnose diseases more accurately.The CAD system extracts,classifies and discriminates features of pulmonary nodules in CT images.The results of detection and recognition are provided as a reference to radiologists and assist them in making a more accurate and objective diagnosis of lung cancer on the basis of reducing part of the volume effect,improving the sensitivity of nodules detection and reducing the false positive rate.At the same time,the three-dimensional structure information provides doctors with a display effect that is closer to a real human body and a qualitative analysis method forspecific diseases,which plays a very important role for medical workers to formulate more accurate medical plans.However,with the increasing volume of data and the complexity of data in three-dimensional data,the artificial feature representation is obviously not the most reasonable choice.Therefore,it has become a goal for researchers to learn the most suitable feature representation by machine learning algorithm.The study of this paper is mainly based on the existing CT image data of the lungs,using stack self-coding technology to achieve the diagnosis of pulmonary nodules and more basic prediction.The main contents and innovations include the two aspects as following:(1)Most of the existing methods of pulmonary nodule diagnosis are based on the existing image information.However it is often combined with the patient’s clinical information to conduct a comprehensive assessment in the actual diagnosis of radiologists.It is lack of clinical information in existing datasets.Based on this,after collecting the relevant clinical information with the cooperative hospital and analyzing and processing,a diagnosis method of pulmonary nodules is proposed.The method combines the clinical multi-classification information with the framework of deep learning.It can perform self-encoding learning after multiple classification of local features on the input CT images.Experiments have shown that stratified analysis of features in CT images is better than single feature diagnosis.Moreover,the method proposed in this paper adds the training sample’s tag information to the training sample in the autoencoder process and uses the semi-supervised method to perform feature learning on the sample.In the traditional sparse self-encoding learning process,some influential factors with a little correlation are removed.Based on this,this article adds important clinical diagnostic information to the network for the treatment of lung diseases.Experimentalresults support that the proposed method is more exact than the traditional sparse autoencoder network.(2)We found that it is difficult to realize the diagnosis of early pulmonary nodule lesions based on CT tomography image-based 2D image processing after a deep study of 2D images.The establishment of three-dimensional assistant diagnosis of visible pulmonary nodules has become the key to early diagnosis of lung cancer.Three-dimensional features can reflect the spatial position and relationship of pulmonary nodules.Three-dimensional features can reflect changes in the size of lung nodules over time,as well as the degree of surface roughness.These characteristics have important guiding value for benign and malignant diagnosis of pulmonary nodules and doctor’s treatment,and it can very useful to improve survival rate.Therefore we mainly use the collected data from hospital and open data sets of lung CT images and their 3D reconstruction images to realize the two-dimensional and multi-view feature learning.Firstly,the three-dimensional image shape is characterized by two-dimensional correlation features.Among them,the thermonuclear features constitute a multi-scale histogram distribution,and at the same time the images are covered in all directions from different multi-viewpoints.Then learn the high-level features from each of the scales using a deep autoencoder network.The feature descriptors are connected to the outputs of multiple scale hidden layers and feature learning is also performed on multi-dimension images.The effectiveness of this method in computer aided pulmonary disease diagnosis is verified by testing on the relevant experimental data set.