The Computer-aided Diagnosis Of Lung Nodule Based On Principal Component Analysis Net And Support Vector Machine

Lung Cancer is the leading cause of cancer death in men and women in China with the development of smog.The prompt and widely available diagnostics of lung cancer are important and urgent.The traditional diagnostics of lung cancer are analyzing X-ray by doctor.The nodule is an important feature for early lung cancer.However,The recognition of nodule requires extensive knowledge and experience of the cytologist.Computer-aided diagnosis can speed up analysis time and provide a second diagnostic.Most CADs are based on pattern recognition.Feature extraction and classifier of nodule images are very important process in pattern recognition.But features in traditional methods are extracted by hands.The shortcoming is accidental and randomness.In this study,features will be learned from a deep learning algorithm.a classifier called SVM which has proved to be an effective classifier in image's recognition will be applied for classifying.Then,this combined method will be applied to classify normal and abnormal lung images.In this paper,we first introduce the process of human brain's automatic feature extraction.The layer and unsupervised learning are the main special characters.So the main characters of deep learning is also layer and unsupervised training.And the algorithms we used in this paper is introduced.At the same time,we introduce the significance of nodules in lung cancer diagnosis.In addition,the distinguish between nodules and normal image is apparent.The deep learning algorithm is suitable for this case.Then,we introduce the implement of three algorithms.Specifically for the widely used picture database called JSTR,PCANet is used to extract the features of nodular images and normal images.SVM is used to classify the two samples.The selection of PCANet filter and some parameters of SVM are the important parts of the implement process.We also post the problems,for example,the long training time and the size of images.These problems may be a reference for other researchers.The next,results are displayed and analyzed.The results show that the PCANet has an excellent performance.The accuracy has 10% improvement compared with no PCANet.And the accuracy is up to 83%.In additional,compared with MacMahon's CAD systems,the accuracy of our algorithms has an improvement of 3 percentage point.compared with Hardie's CAD systems the accuracy of our algorithms has an improvement of 5 percentage point.We can make a conclusion that the PCANet_SVM is effective for the lung CAD.This also can be a reference for the clinical application.At last,we also develop a software in Matlab GUI tool.The main functions of this GUI include reading images,selecting the algorithm for the feature extraction,selecting the kernel of SVM,inputting the parameters for SVM training and reading a new image for prediction.