A Multi-model Fusion Method For Classifying Benign And Malignant Pulmonary Nodules With Embedded Attention Mechanism

Lung cancer is one of the most difficult to cure tumor diseases among all diseases.Early detection and treatment as soon as possible ar e the only effective measures to prevent and treat lung cancer.The main manifestation of lung cancer in the early stage of onset is pulmonary nodules.Doctors can diagnose and treat early lung cancer in a timely manner through the observation and analysis of chest CT.However,with the rapid increase in the number of lung CT images,it is almost impossible to screen and judge from a large number of CT images alone.The computer-aided system based on deep learning can perform preliminary selection and processing of CT images,which can greatly reduce the burden on doctors and reduce the rate of doctors’ misdiagnosis and missed diagnosis.Therefore,the thesis focuses on the application of lung CT images based on deep learning in the computer-aided diagnosis system.The research content includes the segmentation and classification of CT images of lung nodules.In the segmentation network of the traditional U-shaped encoder-decoder structure,the feature fusion part is to restore the size through the feature ma p and then directly splice it.This feature fusion method does not work well in dealing with small target objects with fuzzy edges such as lung nodules.After segmentation,the CT image can remove noise and provide more accurate lung nodule data map for subsequent classification;secondly,different models and different attention mechanisms in deep learning also focus on the feature information of the same image.Different,so fusion of several suitable models can extract more comprehensive feature information for small targets such as lung nodules.This thesis proposes a multi-model fusion network embedded in the attention mechanism to classify lung nodules from benign and malignant.This classification method first segmented the lung nodules to remove the extraneous tissues in the lung CT,and obtained the detailed feature map of the lung nodules,which was used as a branch of the subsequent multi-model input.Then,a multi-model fusion network framework of embedding space and channel attention mechanism is constructed to classify the segmented lung nodules(64×64)and the original size(512×512)lung nodules after preprocessing.This article introduces the classification of benign and malignant lung nodules in CT images in two stages.The first stage is seg mentation of lung nodules;the second stage is the study of benign and malignant lung nodules.In the first stage,for the traditional U-shaped encoder-decoder segmentation model feature fusion,the semantic and abstract feature differences between the hig h and low-level feature map features,directly using upsampling and other operations to restore the size will produce resolution differences and lead to feature fusion effects For the poor problem,a DSF-UNet(Dense SEB Fusion-UNet)model combining dense cross-connection structure and high-level semantic embedding structure is proposed.U-net,which performs well in medical image processing,is selected as the basic framework of the segmentation model,and dense connection modules and advanced semantic embedding modules are introduced in the feature extraction stage,which can efficiently perform cross-layer fusion and take into account the details of lung nodules themselves.information.Through experimental comparison,the proposed lung nodule segmentation method based on feature fusion can accurately segment the lung nodule area,and the Dice similarity coefficient value,accuracy rate and recall rate reached 87.54%,86.93% and 86.67%,respectively.In the second stage,aiming at the problem of the deviati on of the classification effect of lung nodules in CT images due to blurred edges and unobvious features,this thesis proposes a multi-model fusion method embedded in the attention mechanism.In this method,the original CT image is segmented into lung par enchyma and the previous step is segmented to obtain two images of different sizes,which are input into the spatial attention model and the channel attention model for training.Among them,the spatial attention model focuses on extracting lung nodules.F or the spatial location information in CT images,the channel attention model focuses on extracting the detailed features of lung nodules.Finally,the features extracted by the two models are fused to obtain the benign and malignant classification results.A lot of experiments show that this multi-model fusion method can well extract the position information of lung nodules in CT images and their own edge features.Experimental results show that the accuracy,sensitivity,and specificity of the model frame work can reach 95.53%,95.87%,and 95.22%,respectively.Compared with the traditional network model,this method has achieved better classification results.