Application Research On Computer-aided Diagnosis Of Pulmonary Nodules On Thoracic CT Images Based On Deep Learning

With the advent of the era of big data in the 21st century,the field of artificial intelligence represented by deep learning has received widespread attention from scientific research experts,and has shown unprecedented value.Deep learning has been widely used in daily life with its unique characteristics of imitating human brain thinking and abstraction,and to alleviate the pressure and danger of manual labor.Among them,computer aided diagnosis(CAD)in medicine is one of the widely used fields.In this paper,we mainly study the CAD of pulmonary nodules of thoracic CT images in the early stage of lung cancer.Traditional pulmonary nodules of thoracic CT images diagnosis mainly rely on the personal clinical experience of doctors,which has strong subjectivity and low diagnostic accuracy.Therefore,exploring scientific diagnostic methods to detect lung cancer as early-as possible and get effective treatment is the necessary means to improve the survival rate of lung cancer patients.In order to effectively manage the massive CT images and further explore their value,we constructed a CAD medical images knowledge base(KB)of pulmonary nodule to store the case data of chest CT scans.The KB includes two logically independent relational databases,DICOM medical image database and expert diagnostic database.The DICOM medical image database mainly stores medical image information.including CT image,patient information,study information and series information.The expert diagnostic database mainly stores the diagnostic information of pulmonary nodules on CT images,including nodule character information,location information and type information.This medical image KB system is realized based on Apache Web server,and uses PHP scripting language to manage and maintain the MySQL database.At present,the case set of lung cancer stored in this KB mainly comes from LIDC-IDRI database.The realization of the pulmonary nodule CAD medical image KB system facilitates the maintenance and management of pulmonary medical image data and diagnostic information data in the later stage,and preliminarily explores the expression and reasoning of medical image knowledge,and provide reliable labelled pulmonary nodule images for image-based deep learning analysis,and then realize CAD of lung nodules.In order to realize the CAD of pulmonary nodules of thoracic CT images,we adopt image-based deep learning technology to classify pulmonary nodules.The biggest challenge of medical image classification is the scarcity of labeled medical image data.To avoid over-fitting in the training process,we refer to the LeNet and AlexNet models to design a shallow convolutional neural network(CNN)to classify pulmonary nodules,including benign and malignant pulmonary nodules,and different malignancy of'the malianant nodules.The both experiments use the same network structure,so actually these two experiments are combined to be a multilabel classification task.The realization of medical image KB of pulmonary nodules provides reliable labeled images data for the classification of pulmonary nodules.In the training process,we use Holdout method to tune the network structure and parameter settings,and then use 10-fold cross validation method to testify the robustness of the classification model we trained.The result show that LIDCNet model has a good effect on the classification of pulmonary nodules,and the accuracy of Top-1 all more than 95%.