Research And Implementation Of A Convolutional Neural Network Based Lung Nodule Assisted Diagnosis System

Today,lung cancer has become one of the most common fatal malignancies worldwide.Lung nodules are the early stage lesions of lung cancer,and being able to be detected and diagnosed at the early stage is very important for the early treatment of lung cancer.However,the size and shape of lung nodules vary,which makes the diagnosis of lung nodules difficult and requires doctors to judge and analyze them in CT images,which makes the diagnosis of lung nodules very difficult.Computer-aided Detection and Diagnosis(CADx)system is a technology that uses computer technology and image processing technology to assist doctors in diagnosing and detecting medical images.The CADx system can automatically detect abnormal areas in medical images and evaluate patients’ conditions,thus helping doctors to make more accurate diagnoses.In this paper,we design a 3D convolutional neural network-based lung nodule diagnosis system,which uses a two-stage CAD unit + lung nodule classification to complete the diagnosis of lung nodules,and can present the lung nodule detection,segmentation and diagnosis results in various forms to assist doctors in further diagnosis and evaluation of the patient’s condition.Firstly,in the CAD unit,this paper uses a lung nodule diagnosis method that first extracts nodule candidate regions and then segments them,using the YOLOv5 model to detect and extract the regions of interest of lung nodules,reducing the interference of irrelevant regions,reducing the dimensionality and complexity of images,and providing convenience for subsequent segmentation and analysis.Next,by analyzing the existing 3D segmentation model,we found that it has certain limitations,which may lead to inaccurate segmentation and slow segmentation speed.In order to solve this problem,for small target segmentation challenges such as lung nodules,this paper proposes a multi-scale feature fusion MSFFV-Net segmentation network model based on V-Net,redesigns the forward propagation of the network,and uses more efficient activation functions,incorporates depth supervision methods,and attention mechanisms and other modules in the model.Experiments were conducted on the LUNA16 dataset,and the MSFFV-Net segmentation network showed a significant improvement over the original V-Net in different metrics,including a 5%improvement in the DICE coefficient,demonstrating the effectiveness of MSFFV-Net for lung nodule segmentation.And by comparing with other lung nodule segmentation models,MSFFV-Net also obtained better segmentation results.In the lung nodule classification unit,the three main parts include feature extraction,feature selection,and training the classifier.Firstly,the segmentation result images in the CAD unit are feature extracted using Py Radiomics,and the key features are selected using the feature selection algorithm to reduce the influence of redundant features and improve the performance of the classifier.Finally,different classifiers are trained to achieve the classification of lung nodules,and finally the SVM with better classification performance is selected as the classifier of this system.In this paper,a desktop application for lung nodule assisted diagnosis system is created based on the above algorithm by using Py Qt5 toolkit.The application provides a simple and easy-to-understand user interface,including lung nodule detection,segmentation and classification functions.Physicians can upload medical imaging data and analyze it through simple operations.The system will automatically perform lung nodule detection,segmentation and classification and present the results to the physician in graphical and tabular form.The system was developed to improve the accuracy and efficiency of pulmonary nodule diagnosis,provide doctors with more convenient and reliable pulmonary nodule diagnosis services,and also provide patients with a safer and more comfortable diagnosis experience.