Lung Nodules Diagnosis Using Deep Auto-encoder Based On Generalized Extreme Learning Machine

Lung cancer is the leading cause of death worldwide.Meanwhile early diagnosis and treatment of lung cancer is the most effective way to improve the survival rate of patients.Lung nodules are early manifestations of lung cancer.With the development of computed tomography imaging,CT becomes the standard way to detect pulmonary nodules,but further diagnosis of lung nodules requires further pathological examination.As a result,millions of medical CT images are accumulated and require further diagnosis by a radiologist.Facing of massive lung CT images,the computer-aided diagnosis system can effectively assist doctors in the diagnosis of pulmonary nodules,thereby reducing the workload of doctors processing data.However,the traditional lung noduleassisted diagnosis system based on manual extraction features(texture features,shape features,grayscale features)is not suitable for the complex lung nodule structure.In recent years,deep learning methods have been applied to the diagnosis of lung nodules.In recent years,the application of deep learning methods to the diagnosis of pulmonary nodules has become a trend.However,in the traditional deep learning-based pulmonary nodule diagnosis method,there is a problem that only the deep learning is used as a feature extraction tool to extract image features and ignore the prior knowledge to play a role in the diagnosis of pulmonary nodules.Therefore,how to integrate prior knowledge into the deep model-based pulmonary nodule diagnosis system to further improve the diagnostic accuracy of lung nodules becomes very meaningful At the same time,facing of the problem of multi-training speed of most deep learning models,how to improve the training speed while ensuring accuracy is also the focus of this paper.The deep auto-encoder technique can classify and diagnose lung nodules by learning the features in multi-layer auto-encoder stacked end-to-end lung CT images and using the learned features.Based on the deep auto-encoder,this paper proposes a deep auto-encoder based on generalized extreme learning machine for the feature extraction ability training speed and feature robustness problem.The study used deep auto-encoder based on generalized extreme learning machine technology to achieve benign and malignant auxiliary diagnosis of lung nodules.The main research contents and innovations of this paper are as follows:(1)Because medical images are difficult to collect and the standards are different,the amount of medical images is relatively small compared with the amount of data in other fields,which hinders the development of medical deep learning.The amount of data determines whether the deep network can learn the complex features of lung nodules,so solving this problem has important significance for the classification and diagnosis of lung nodules.Based on this,in order to solve the problem that the classification accuracy is not high due to the insufficient comprehensive characteristics of the learned pulmonary nodules,a multi-feature based deep auto-encoder based on generalized extreme learning machine.First,three data sets of lung nodule images are constructed as input by image preprocessing;Then,the manifold learning is integrated into the deep autoencoder based on the extreme learning machine to form unsupervised generalized deep auto-encoder.Finally,the fusion eigenvectors are obtained through different fusion strategies,and the lung nodules are diagnosed by benign and malignant classification by classifier.The experimental results show that the prior knowledge can be used as a guide to improve the diagnostic accuracy,sensitivity and specificity of pulmonary nodules compared with single image input.At the same time,the deep auto-encoder based on generalized extreme learning machine network formed by manifold learning has better feature extraction ability than the original network model.(2)The bilinear model can be effectively used for feature representation.In order to further improve the stability and feature extraction ability of deep auto-encoder based on generalized extreme learning machine,the performance instability problem usually caused by AE-ELM random input weight is overcome.In this paper,deep auto-encoder based on generalized extreme learning machine is extended to bilinear structure to further improve the stability and feature extraction ability of the network structure.Then,the high-dimensional bilinear feature generated by the bilinear structure is reduced by the asymmetric dimensionality reduction strategy.Finally,the weighted voting method in integrated learning is used to fuse the reduced bilinear features to obtain the final diagnosis results of lung nodules.Experiments show that the generalized deep self-encoding extended to bilinear has more robust and efficient feature extraction ability,and experiments show that the method is effective in the diagnosis of pulmonary nodules.