Research On Deep Learning-based Segmentation And Classification Of Benign And Malignant Pulmonary Nodules In CT

Among all cancers,lung cancer has the highest incidence and mortality rate,which is a serious threat to people’s life and health.Early screening of lung cancer by computed tomography(CT)can reduce the mortality rate of lung cancer.Lung nodules are one of the major manifestations of lung cancer,but the characteristics of lung nodules on CT images are similar to other tissues in the lung parenchyma and the screening process requires a high level of attention and expertise,making it easy for less experienced physicians to miss and misdiagnose.The use of deep learning algorithms to assist in diagnosis can avoid these situations.For this reason,it is important to study deep learning-based lung nodule segmentation and classification to achieve stability,accuracy and efficiency of computer-aided diagnosis system.To address the problems of noise interference and class-polar imbalance in lung nodule segmentation data,the lung parenchyma and region of interest are extracted using traditional image algorithms.For the problems of class imbalance and insufficient data volume in benign and malignant lung nodule classification data,Wasserstein generative adversarial network with conditions and data enhancement methods are used.To address the problem that U-Net cannot accurately segment lung nodules,especially small-sized nodules,a REMU-Net-based lung nodule segmentation method is proposed.Res Ne St,which introduces the spatial attention module,is used as the encoder of U-Net to improve the diversity and effectiveness of the network extracted features.The enhancement module is proposed based on the atrous spatial pyramid pooling and used as the transition structure of U-Net to make full use of the contextual information of deep features.The skip connection of U-Net is replaced with multi-scale skip connection to avoid the limitation that the decoder subnet can only accept same-scale information.The improved algorithm improves Dice by 6.65% to84.76% and has better segmentation performance than other improved U-Net methods.To address the problems of low accuracy and flexibility of existing classification methods,DCA-Xception is proposed for the classification of benign and malignant lung nodules.An adaptive two-channel feature extraction module is introduced in Xception to obtain information around the target,and a convolutional block attention module is introduced to enable the network to learn features in a targeted manner.Experiments show that the improved network has an accuracy of 83.46% and an AUC of 92.89%,which is 4.58% and 3.52% higher than the original model,respectively,and is able to achieve an accuracy of 85.24% on K-nearest neighbor and random forest classifiers if the network is utilized as a feature extractor.DCA-Xception has better classification performance compared with classical classification networks and fine-tuned pre-trained networks.