Lung Data Processing Based On Non-negative Matrix Factorization

Accurate diagnosis of pulmonary diseases and effective treatment are the guarantees of early recovery of patients.In clinical practice,the main methods of lung system diagnosis are chest X-ray,CT,lung sound auscultation,etc.Therefore,the forms of lung data are diverse,including lung sound and lung image data.Lung sound signal contains more comprehensive pathological information of lung system.Lung sound auscultation has the advantages of convenience and high diagnostic efficiency,which is widely used in the clinical diagnosis of the lung.But in the process of auscultation of lung sound,it is easy to be interfered by heart sound,which affects the doctor’s diagnosis.Therefore,it is of practical significance to separate pure lung sounds from cardiopulmonary sounds.In addition,lung diseases such as lung cancer need to be diagnosed by lung image data.At present,most computer-aided pathological diagnoses of lung cancer are using artificial intelligence technology to detect the CT image of the lung in the early stage of operation.However,the rapid pathological classification of pulmonary nodules during surgery can help to enhance doctors’ judgment of the disease,and it is very important for patients to choose the follow-up treatment plan.Non-negative matrix factorization(NMF)performs well in dimensionality reduction and feature extraction.This study is based on NMF technology,it is mainly to solve two tasks of cardiopulmonary sound separation and lung cancer pathological image classification.In the aspect of cardiopulmonary sound separation,based on Sparse Dual Graphregularized non-negative matrix factorization with L21 norm constraint(L21SDGNMF),this paper improves the traditional NMF based cardiopulmonary sounds separation method in three stages: decomposition,clustering,and reconstruction.L21 SDGNMF is used to decompose the mixed time-frequency matrix,L21 SDGANMF with label constraint is used to obtain the reference signal,and L21 SDGNMF is used to update the coefficient matrix.Through simulation experiments,the effectiveness of the improved method based on L21 SDGNMF is verified,and the signal-to-noise ratio and similarity are significantly improved compared with the traditional method based on NMF.In the pathological diagnosis of lung cancer,we first expand the samples of lung cancer pathological image data during operation and make data sets with different numbers of samples for experimental verification.In addition,the color histogram feature extraction,nonnegative matrix factorization,and support vector machine classification are combined to classify the pathological images of lung cancer according to the two tasks of distinguishing benign from malignant and invasive cancer.In the simulation experiment,compared with the deep neural network RESNET18 and SVM classification method,the COLORNMFSVM method proposed in this paper has higher classification accuracy,and can effectively classify the lung cancer pathological image.