Research On Detection And Classification Of Lung Nodules Based On Deep Learning

Lung cancer is a type of cancer with high morbidity and high mortality worldwide,which has caused great harm to human physical and mental health.Early screening and treatment of lung cancer are essential to improve the survival rate and quality of life of patients.Pulmonary nodule lesions are an important pathological characterization of lung cancer in the early stage of development.As a non-invasive imaging tool,Computed tomography(CT)can accurately scan and image human lung tissues and lung nodular lesions while reducing patient pain.The use of deep learning technology to achieve accurate detection and classification of benign and malignant pulmonary nodules in lung CT images is of significant for the early prevention and treatment of lung cancer.The research on computer-aided detection and classification of benign and malignant pulmonary nodules still faces many challenges.On the one hand,the tissue structure in the lung is complex,and there are many other tissues or organs with similar visual characteristics to lung nodules,which affect the detection effect of the algorithm.On the other hand,pulmonary nodules have the characteristics of small image differences between different types and significant image differences within the same type in image features,which makes it challenging to classify benign and malignant.To address the above problems,this thesis proposed methods for pulmonary nodules detection and benign and malignant classification based on deep learning technology.The main research contents of this thesis can be detailed as follows:(1)Aiming at the problems of small nodule size and many interfering tissues in the lung cavity,a pulmonary nodule detection algorithm is proposed based on a threedimensional convolutional neural network.Firstly,the 2D U-Net++structure is extended to 3D,and its encoder,decoder structure,and internal nested connection structure are retained,and the effective fusion of image shallow and deep features is realized to adapt to the detection of nodules of different sizes.Second,a cross-stage partial fusion module is added to the feature extraction network to increase the feature learning ability of the network and reduce feature redundancy.Finally,the 3D region generation module is used to output the location information directly and predict the probability of nodules to achieve end-to-end detection of pulmonary nodules.The experimental results of cross-validation on the LUNA16 data set show that the CPM score of the proposed method reaches 86.2%,and the detection sensitivity reaches 95.4%when the average number of false positives is 8.0,and it gets good detection sensitivity for nodules in all sizes.(2)Aiming at the similar image features of benign and malignant pulmonary nodules,a benign and malignant pulmonary nodules classification method based on Transformer structure and residual network is proposed.In order to adapt to the classification of pulmonary nodules of different types and sizes,the remote dependence relationship between image pixels is established by designing local feature and global feature extraction modules.Specifically,residual blocks with convolution operations are used to extract local features,while Transformer feature extraction modules incorporating the attention mechanism are used to capture global features.In addition,a sequence fusion module is designed to aggregate and extract the sequence feature information output by the Transformer module to improve the classification accuracy.The results of ten-fold cross-validation on the LDC-IDRI dataset show that the proposed method achieved 96.28%AUC value and 92.92%accuracy and has good performance in benign and malignant classification of pulmonary nodules.