Research Of Solitary Pulmonary Nodules Diagnosis Method Based On Deep Auto-encoder

In recent years,lung cancer has become one of highest incidence and mortality of cancer.Early diagnosis and treatment of lung cancer can effectively reduce the mortality of patients,and plays a vital role in improving the survival rate.The early stage of lung cancer are solitary pulmonary nodules.The radiologist could effectively detect the solitary pulmonary nodules through PETCT scan.However,the diagnosis of benign or malignant pulmonary nodules needs further pathological examination,which brings economic and mental burden to patients.At present,computer-aided diagnosis technology can provide support for the radiologist diagnosis through extracting the statistical feature in pulmonary nodule images and analyzing the images based on the extracted feature set.However,diagnostic methods based on statistical feature extraction cannot be applied to complex pulmonary nodules with complex organizational structure.The deep auto-encoder technology can capture image features automatically through multilayer nonlinear network stacking,and classify the images according to these features,which objectively reflects the real information of the image.Therefore,this study mainly uses the deep auto-encoder to realize the diagnosis of benign and malignant pulmonary nodules.The major work of this paper includes two parts below:(1)The existing methods of diagnosis of pulmonary nodules mainly depend on much prior knowledge such as image processing techniques and professional diagnostic knowledge to extract image features.However,these manual features probably are unsuitable for all kinds of pulmonary nodule images.In order to address this problem,an automatic diagnosis method for pulmonary nodule images based on the stacked extreme learning machine(ELM)was proposed in this study.As the systematic noise on CT imaging would directly lead to low contrast,an adaptive histogram equalization method was firstly used in this study to highlight the contrast of original pulmonary nodule images.Then,the enhanced images were input into an ELM-based auto-encoder to learn the unsupervised feature of pulmonary nodule images.Finally,the high-level features from the deep learning network were used as the input of the supervised original ELM to realize the final diagnosis.Because the extreme learning machine algorithm does not require parameter tuning,the training speed of the ELM-based auto-encoder is faster than that of the BP-based auto-encoder,and the diagnosis performance of ELM network is similar to the BP network.Compared with the existing diagnostic methods,the method in this study was a reliable for diagnosis of pulmonary nodules in terms of accuracy,sensitivity and specificity.(2)In order to overcome the one-sidedness of single modal image feature and the limitation of unsupervised feature learning,a novel pulmonary nodule diagnosis method was proposed in this study using dual-modal deep supervised auto-encoder based on extreme learning machine.In this method,a deep supervised denoising auto-encoder(SDAE)was introduced to learn discriminative features automatically from Computed Tomography(CT)and Positron Emission Tomography(PET)images.The network was fed with nodule images in pairs obtained from CT and PET respectively.For each pair images,the high level discriminative features of nodules in CT and PET were extracted from stacked supervised auto-encoder layers.The outputs of the proposed architecture were combined using different fusion methods to get the final classification.The experimental results showed that the fused features had more discrimination than individual features.Moreover,extensive experiments showed that the method obtained an accuracy of 92.81±0.57% and 1.58 false positive per scan and was robust against noisy input,which outperformed other state-of-the-art methods in pulmonary nodule diagnosis.It has certain reference value for the further study of pulmonary nodules diagnosis.